[
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 1024,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 325.890395,
    "ttft_ms_q25": 311.19831550000004,
    "ttft_ms_q75": 326.66981050000004,
    "prefill_tokens_per_s": 3093.0644642042917,
    "prefill_tokens_per_s_q25": 3085.7021551973376,
    "prefill_tokens_per_s_q75": 3246.32784786788,
    "handoff_plus_first_decode_ms": 50.954615000000004,
    "handoff_plus_first_decode_ms_q25": 50.2046335,
    "handoff_plus_first_decode_ms_q75": 53.7457075,
    "steady_tpot_ms": 38.1771325,
    "steady_tpot_ms_q25": 37.48788975,
    "steady_tpot_ms_q75": 38.18753075,
    "steady_tpot_p95_ms": 40.4363914,
    "steady_tpot_p95_ms_q25": 39.8993059,
    "steady_tpot_p95_ms_q75": 41.703343125,
    "decode_tokens_per_s": 25.393661298258756,
    "decode_tokens_per_s_q25": 25.317242057478495,
    "decode_tokens_per_s_q75": 25.67713548203487,
    "decode_ms": 590.698593,
    "decode_ms_q25": 584.248521,
    "decode_ms_q75": 592.486995,
    "end_to_end_ms": 916.588988,
    "end_to_end_ms_q25": 895.4468365,
    "end_to_end_ms_q75": 919.1568055,
    "output_tokens_per_s_e2e": 17.456024684424857,
    "output_tokens_per_s_e2e_q25": 17.407394231742355,
    "output_tokens_per_s_e2e_q75": 17.878140595271027,
    "all_tokens_per_s_e2e": 1117.1855798031909,
    "all_tokens_per_s_e2e_q25": 1114.0732308315107,
    "all_tokens_per_s_e2e_q75": 1144.2009980973457,
    "energy_net_psys_j": 28.611885485606145,
    "energy_net_psys_j_q25": 28.54256725633332,
    "energy_net_psys_j_q75": 28.867325691640882,
    "prefill_energy_net_psys_j": 9.723831623047879,
    "prefill_energy_net_psys_j_q25": 9.373954739550998,
    "prefill_energy_net_psys_j_q75": 9.876530738910652,
    "decode_energy_net_psys_j": 19.09353604290219,
    "decode_energy_net_psys_j_q25": 18.921476723457403,
    "decode_energy_net_psys_j_q75": 19.34067183622711,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 1024,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 325.890395,
    "ttft_ms_q25": 311.19831550000004,
    "ttft_ms_q75": 326.66981050000004,
    "prefill_tokens_per_s": 3093.0644642042917,
    "prefill_tokens_per_s_q25": 3085.7021551973376,
    "prefill_tokens_per_s_q75": 3246.32784786788,
    "handoff_plus_first_decode_ms": 44.62128,
    "handoff_plus_first_decode_ms_q25": 43.1801255,
    "handoff_plus_first_decode_ms_q75": 75.318545,
    "steady_tpot_ms": 37.889624,
    "steady_tpot_ms_q25": 37.802927249999996,
    "steady_tpot_ms_q75": 37.9378635,
    "steady_tpot_p95_ms": 38.696261199999995,
    "steady_tpot_p95_ms_q25": 38.664764500000004,
    "steady_tpot_p95_ms_q75": 38.79098082499999,
    "decode_tokens_per_s": 25.956409774464063,
    "decode_tokens_per_s_q25": 24.74549161705678,
    "decode_tokens_per_s_q75": 26.05161174600982,
    "decode_ms": 577.89194,
    "decode_ms_q25": 575.787804,
    "decode_ms_q75": 607.6260605,
    "end_to_end_ms": 903.782335,
    "end_to_end_ms_q25": 902.4576145000001,
    "end_to_end_ms_q75": 918.824376,
    "output_tokens_per_s_e2e": 17.70337766117104,
    "output_tokens_per_s_e2e_q25": 17.41822453030541,
    "output_tokens_per_s_e2e_q75": 17.729402709710655,
    "all_tokens_per_s_e2e": 1133.0161703149465,
    "all_tokens_per_s_e2e_q25": 1114.7663699395462,
    "all_tokens_per_s_e2e_q75": 1134.681773421482,
    "energy_net_psys_j": 23.784934389537618,
    "energy_net_psys_j_q25": 23.55439354148898,
    "energy_net_psys_j_q75": 24.002253128439733,
    "prefill_energy_net_psys_j": 9.723831623047879,
    "prefill_energy_net_psys_j_q25": 9.373954739550998,
    "prefill_energy_net_psys_j_q75": 9.876530738910652,
    "decode_energy_net_psys_j": 13.75570453476419,
    "decode_energy_net_psys_j_q25": 13.677862802578328,
    "decode_energy_net_psys_j_q75": 14.47559927302596,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0141700634146606,
    "energy_reduction_vs_gpu": 0.16870440427621713
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 1024,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 372.282625,
    "ttft_ms_q25": 371.818852,
    "ttft_ms_q75": 373.5173075,
    "prefill_tokens_per_s": 2707.6203193742926,
    "prefill_tokens_per_s_q25": 2698.6996157778267,
    "prefill_tokens_per_s_q75": 2711.0017762854477,
    "handoff_plus_first_decode_ms": 44.62128,
    "handoff_plus_first_decode_ms_q25": 43.1801255,
    "handoff_plus_first_decode_ms_q75": 75.318545,
    "steady_tpot_ms": 37.889624,
    "steady_tpot_ms_q25": 37.802927249999996,
    "steady_tpot_ms_q75": 37.9378635,
    "steady_tpot_p95_ms": 38.696261199999995,
    "steady_tpot_p95_ms_q25": 38.664764500000004,
    "steady_tpot_p95_ms_q75": 38.79098082499999,
    "decode_tokens_per_s": 25.956409774464063,
    "decode_tokens_per_s_q25": 24.74549161705678,
    "decode_tokens_per_s_q75": 26.05161174600982,
    "decode_ms": 577.89194,
    "decode_ms_q25": 575.787804,
    "decode_ms_q75": 607.6260605,
    "end_to_end_ms": 950.174565,
    "end_to_end_ms_q25": 949.3051115000001,
    "end_to_end_ms_q75": 979.4449125,
    "output_tokens_per_s_e2e": 16.839011050564164,
    "output_tokens_per_s_e2e_q25": 16.35038584129911,
    "output_tokens_per_s_e2e_q75": 16.85444777204438,
    "all_tokens_per_s_e2e": 1077.6967072361065,
    "all_tokens_per_s_e2e_q25": 1046.4246938431431,
    "all_tokens_per_s_e2e_q75": 1078.6846574108404,
    "energy_net_psys_j": 20.752796211965624,
    "energy_net_psys_j_q25": 20.599269091711488,
    "energy_net_psys_j_q75": 21.559474802474096,
    "prefill_energy_net_psys_j": 6.9970916772014355,
    "prefill_energy_net_psys_j_q25": 6.92140628913316,
    "prefill_energy_net_psys_j_q75": 7.083875529448138,
    "decode_energy_net_psys_j": 13.75570453476419,
    "decode_energy_net_psys_j_q25": 13.677862802578328,
    "decode_energy_net_psys_j_q75": 14.47559927302596,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.9646532561098391,
    "energy_reduction_vs_gpu": 0.2746791810555167
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 1024,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 372.282625,
    "ttft_ms_q25": 371.818852,
    "ttft_ms_q75": 373.5173075,
    "prefill_tokens_per_s": 2707.6203193742926,
    "prefill_tokens_per_s_q25": 2698.6996157778267,
    "prefill_tokens_per_s_q75": 2711.0017762854477,
    "handoff_plus_first_decode_ms": 50.954615000000004,
    "handoff_plus_first_decode_ms_q25": 50.2046335,
    "handoff_plus_first_decode_ms_q75": 53.7457075,
    "steady_tpot_ms": 38.1771325,
    "steady_tpot_ms_q25": 37.48788975,
    "steady_tpot_ms_q75": 38.18753075,
    "steady_tpot_p95_ms": 40.4363914,
    "steady_tpot_p95_ms_q25": 39.8993059,
    "steady_tpot_p95_ms_q75": 41.703343125,
    "decode_tokens_per_s": 25.393661298258756,
    "decode_tokens_per_s_q25": 25.317242057478495,
    "decode_tokens_per_s_q75": 25.67713548203487,
    "decode_ms": 590.698593,
    "decode_ms_q25": 584.248521,
    "decode_ms_q75": 592.486995,
    "end_to_end_ms": 962.9812179999999,
    "end_to_end_ms_q25": 956.067373,
    "end_to_end_ms_q75": 966.0043025,
    "output_tokens_per_s_e2e": 16.61506964095327,
    "output_tokens_per_s_e2e_q25": 16.56323544083485,
    "output_tokens_per_s_e2e_q75": 16.736097496587764,
    "all_tokens_per_s_e2e": 1063.3644570210092,
    "all_tokens_per_s_e2e_q25": 1060.0470682134303,
    "all_tokens_per_s_e2e_q75": 1071.1102397816169,
    "energy_net_psys_j": 26.090627720103626,
    "energy_net_psys_j_q25": 25.842883012590566,
    "energy_net_psys_j_q75": 26.424547365675245,
    "prefill_energy_net_psys_j": 6.9970916772014355,
    "prefill_energy_net_psys_j_q25": 6.92140628913316,
    "prefill_energy_net_psys_j_q75": 7.083875529448138,
    "decode_energy_net_psys_j": 19.09353604290219,
    "decode_energy_net_psys_j_q25": 18.921476723457403,
    "decode_energy_net_psys_j_q75": 19.34067183622711,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9518243667344299,
    "energy_reduction_vs_gpu": 0.0881192456460409
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 2048,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 675.668286,
    "ttft_ms_q25": 659.165996,
    "ttft_ms_q75": 682.1660465,
    "prefill_tokens_per_s": 3007.3928910139794,
    "prefill_tokens_per_s_q25": 2979.017189348549,
    "prefill_tokens_per_s_q75": 3084.6165916339287,
    "handoff_plus_first_decode_ms": 52.48689,
    "handoff_plus_first_decode_ms_q25": 51.409315,
    "handoff_plus_first_decode_ms_q75": 55.5279295,
    "steady_tpot_ms": 39.0188285,
    "steady_tpot_ms_q25": 38.6505935,
    "steady_tpot_ms_q75": 39.2626815,
    "steady_tpot_p95_ms": 42.16395955,
    "steady_tpot_p95_ms_q25": 40.912344225,
    "steady_tpot_p95_ms_q75": 42.25459372500001,
    "decode_tokens_per_s": 24.565094490356536,
    "decode_tokens_per_s_q25": 24.533837957858978,
    "decode_tokens_per_s_q75": 25.002701786304478,
    "decode_ms": 610.622524,
    "decode_ms_q25": 600.119001,
    "decode_ms_q75": 611.4014599999998,
    "end_to_end_ms": 1278.279285,
    "end_to_end_ms_q25": 1266.5616934999998,
    "end_to_end_ms_q75": 1282.2850475,
    "output_tokens_per_s_e2e": 12.516826477400047,
    "output_tokens_per_s_e2e_q25": 12.477846618621317,
    "output_tokens_per_s_e2e_q75": 12.63370717939705,
    "all_tokens_per_s_e2e": 1602.153789107206,
    "all_tokens_per_s_e2e_q25": 1597.1643671835286,
    "all_tokens_per_s_e2e_q75": 1617.1145189628223,
    "energy_net_psys_j": 38.90379185876414,
    "energy_net_psys_j_q25": 38.74473017794784,
    "energy_net_psys_j_q75": 38.95703506467657,
    "prefill_energy_net_psys_j": 20.457327416992882,
    "prefill_energy_net_psys_j_q25": 20.131066499921666,
    "prefill_energy_net_psys_j_q75": 20.46961047374662,
    "decode_energy_net_psys_j": 18.42189832826378,
    "decode_energy_net_psys_j_q25": 18.275119704201217,
    "decode_energy_net_psys_j_q75": 18.813685508001164,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 2048,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 675.668286,
    "ttft_ms_q25": 659.165996,
    "ttft_ms_q75": 682.1660465,
    "prefill_tokens_per_s": 3007.3928910139794,
    "prefill_tokens_per_s_q25": 2979.017189348549,
    "prefill_tokens_per_s_q75": 3084.6165916339287,
    "handoff_plus_first_decode_ms": 74.54955100000001,
    "handoff_plus_first_decode_ms_q25": 74.254223,
    "handoff_plus_first_decode_ms_q75": 93.40781150000001,
    "steady_tpot_ms": 39.836130999999995,
    "steady_tpot_ms_q25": 39.81991425,
    "steady_tpot_ms_q75": 39.872106249999995,
    "steady_tpot_p95_ms": 40.625068549999995,
    "steady_tpot_p95_ms_q25": 40.58012035,
    "steady_tpot_p95_ms_q75": 41.282516625,
    "decode_tokens_per_s": 23.642482614634385,
    "decode_tokens_per_s_q25": 22.952130755722422,
    "decode_tokens_per_s_q75": 23.663828295749074,
    "decode_ms": 634.451138,
    "decode_ms_q25": 633.8793544999999,
    "decode_ms_q75": 654.1258695,
    "end_to_end_ms": 1316.4643070000002,
    "end_to_end_ms_q25": 1312.720082,
    "end_to_end_ms_q75": 1319.789626,
    "output_tokens_per_s_e2e": 12.153766657343942,
    "output_tokens_per_s_e2e_q25": 12.123221199944373,
    "output_tokens_per_s_e2e_q75": 12.188531574297329,
    "all_tokens_per_s_e2e": 1555.6821321400246,
    "all_tokens_per_s_e2e_q25": 1551.7723135928798,
    "all_tokens_per_s_e2e_q75": 1560.1320415100581,
    "energy_net_psys_j": 35.82245416763587,
    "energy_net_psys_j_q25": 35.669914634341026,
    "energy_net_psys_j_q75": 35.91243279987999,
    "prefill_energy_net_psys_j": 20.457327416992882,
    "prefill_energy_net_psys_j_q25": 20.131066499921666,
    "prefill_energy_net_psys_j_q75": 20.46961047374662,
    "decode_energy_net_psys_j": 15.340560637135509,
    "decode_energy_net_psys_j_q25": 15.200304160594403,
    "decode_energy_net_psys_j_q75": 15.76908324320458,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9709942595504033,
    "energy_reduction_vs_gpu": 0.07920404525899993
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 2048,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 703.072141,
    "ttft_ms_q25": 702.349346,
    "ttft_ms_q75": 703.605942,
    "prefill_tokens_per_s": 2890.172830784942,
    "prefill_tokens_per_s_q25": 2887.981820885866,
    "prefill_tokens_per_s_q75": 2893.150201678689,
    "handoff_plus_first_decode_ms": 74.54955100000001,
    "handoff_plus_first_decode_ms_q25": 74.254223,
    "handoff_plus_first_decode_ms_q75": 93.40781150000001,
    "steady_tpot_ms": 39.836130999999995,
    "steady_tpot_ms_q25": 39.81991425,
    "steady_tpot_ms_q75": 39.872106249999995,
    "steady_tpot_p95_ms": 40.625068549999995,
    "steady_tpot_p95_ms_q25": 40.58012035,
    "steady_tpot_p95_ms_q75": 41.282516625,
    "decode_tokens_per_s": 23.642482614634385,
    "decode_tokens_per_s_q25": 22.952130755722422,
    "decode_tokens_per_s_q75": 23.663828295749074,
    "decode_ms": 634.451138,
    "decode_ms_q25": 633.8793544999999,
    "decode_ms_q75": 654.1258695,
    "end_to_end_ms": 1337.523279,
    "end_to_end_ms_q25": 1336.2287004999998,
    "end_to_end_ms_q75": 1357.7318115,
    "output_tokens_per_s_e2e": 11.962408618384877,
    "output_tokens_per_s_e2e_q25": 11.786970890332615,
    "output_tokens_per_s_e2e_q75": 11.97400939834125,
    "all_tokens_per_s_e2e": 1531.1883031532643,
    "all_tokens_per_s_e2e_q25": 1508.7322739625747,
    "all_tokens_per_s_e2e_q75": 1532.67320298768,
    "energy_net_psys_j": 31.684558524914166,
    "energy_net_psys_j_q25": 31.212141183883652,
    "energy_net_psys_j_q75": 31.86194934984863,
    "prefill_energy_net_psys_j": 15.841734325509437,
    "prefill_energy_net_psys_j_q25": 15.760705242154641,
    "prefill_energy_net_psys_j_q75": 16.092866106644045,
    "decode_energy_net_psys_j": 15.340560637135509,
    "decode_energy_net_psys_j_q25": 15.200304160594403,
    "decode_energy_net_psys_j_q75": 15.76908324320458,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.9557061959741787,
    "energy_reduction_vs_gpu": 0.18556631600484064
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 2048,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 703.072141,
    "ttft_ms_q25": 702.349346,
    "ttft_ms_q75": 703.605942,
    "prefill_tokens_per_s": 2890.172830784942,
    "prefill_tokens_per_s_q25": 2887.981820885866,
    "prefill_tokens_per_s_q75": 2893.150201678689,
    "handoff_plus_first_decode_ms": 52.48689,
    "handoff_plus_first_decode_ms_q25": 51.409315,
    "handoff_plus_first_decode_ms_q75": 55.5279295,
    "steady_tpot_ms": 39.0188285,
    "steady_tpot_ms_q25": 38.6505935,
    "steady_tpot_ms_q75": 39.2626815,
    "steady_tpot_p95_ms": 42.16395955,
    "steady_tpot_p95_ms_q25": 40.912344225,
    "steady_tpot_p95_ms_q75": 42.25459372500001,
    "decode_tokens_per_s": 24.565094490356536,
    "decode_tokens_per_s_q25": 24.533837957858978,
    "decode_tokens_per_s_q75": 25.002701786304478,
    "decode_ms": 610.622524,
    "decode_ms_q25": 600.119001,
    "decode_ms_q75": 611.4014599999998,
    "end_to_end_ms": 1312.2490750000002,
    "end_to_end_ms_q25": 1302.4683470000002,
    "end_to_end_ms_q75": 1314.284607,
    "output_tokens_per_s_e2e": 12.192807222973274,
    "output_tokens_per_s_e2e_q25": 12.173952501741823,
    "output_tokens_per_s_e2e_q75": 12.285060393229974,
    "all_tokens_per_s_e2e": 1560.679324540579,
    "all_tokens_per_s_e2e_q25": 1558.2659202229534,
    "all_tokens_per_s_e2e_q75": 1572.4877303334367,
    "energy_net_psys_j": 34.765896216042435,
    "energy_net_psys_j_q25": 34.28695672749047,
    "energy_net_psys_j_q75": 34.90655161464521,
    "prefill_energy_net_psys_j": 15.841734325509437,
    "prefill_energy_net_psys_j_q25": 15.760705242154641,
    "prefill_energy_net_psys_j_q75": 16.092866106644045,
    "decode_energy_net_psys_j": 18.42189832826378,
    "decode_energy_net_psys_j_q25": 18.275119704201217,
    "decode_energy_net_psys_j_q75": 18.813685508001164,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9741133061953196,
    "energy_reduction_vs_gpu": 0.10636227074584059
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 4096,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1709.174643,
    "ttft_ms_q25": 1674.974679,
    "ttft_ms_q75": 1711.6359975,
    "prefill_tokens_per_s": 2387.117089941522,
    "prefill_tokens_per_s_q25": 2383.689314159418,
    "prefill_tokens_per_s_q75": 2436.8736578563226,
    "handoff_plus_first_decode_ms": 60.345018,
    "handoff_plus_first_decode_ms_q25": 60.0690785,
    "handoff_plus_first_decode_ms_q75": 63.59227800000001,
    "steady_tpot_ms": 39.9686285,
    "steady_tpot_ms_q25": 39.96859825,
    "steady_tpot_ms_q75": 40.370734,
    "steady_tpot_p95_ms": 43.72544909999999,
    "steady_tpot_p95_ms_q25": 43.60656337499999,
    "steady_tpot_p95_ms_q75": 43.8626747,
    "decode_tokens_per_s": 23.812829937474554,
    "decode_tokens_per_s_q25": 23.522170856323314,
    "decode_tokens_per_s_q75": 23.832966702447706,
    "decode_ms": 629.9125320000005,
    "decode_ms_q25": 629.3807605000004,
    "decode_ms_q75": 637.7936290000004,
    "end_to_end_ms": 2339.0871750000006,
    "end_to_end_ms_q25": 2312.7683080000006,
    "end_to_end_ms_q75": 2341.0167580000007,
    "output_tokens_per_s_e2e": 6.840275202654642,
    "output_tokens_per_s_e2e_q25": 6.83464174937601,
    "output_tokens_per_s_e2e_q75": 6.9190122567411425,
    "all_tokens_per_s_e2e": 1751.1104518795883,
    "all_tokens_per_s_e2e_q25": 1749.6682878402585,
    "all_tokens_per_s_e2e_q75": 1771.2671377257325,
    "energy_net_psys_j": 72.45499675437996,
    "energy_net_psys_j_q25": 72.0342799550997,
    "energy_net_psys_j_q75": 72.83194131203624,
    "prefill_energy_net_psys_j": 51.94530337567546,
    "prefill_energy_net_psys_j_q25": 51.755696855872586,
    "prefill_energy_net_psys_j_q75": 52.496566841306276,
    "decode_energy_net_psys_j": 20.1610555627554,
    "decode_energy_net_psys_j_q25": 19.914657671449696,
    "decode_energy_net_psys_j_q75": 20.52498099053282,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 4096,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1709.174643,
    "ttft_ms_q25": 1674.974679,
    "ttft_ms_q75": 1711.6359975,
    "prefill_tokens_per_s": 2387.117089941522,
    "prefill_tokens_per_s_q25": 2383.689314159418,
    "prefill_tokens_per_s_q75": 2436.8736578563226,
    "handoff_plus_first_decode_ms": 116.282667,
    "handoff_plus_first_decode_ms_q25": 113.9857415,
    "handoff_plus_first_decode_ms_q75": 132.1389845,
    "steady_tpot_ms": 42.9531705,
    "steady_tpot_ms_q25": 42.88008375,
    "steady_tpot_ms_q75": 43.00171325,
    "steady_tpot_p95_ms": 44.112806049999996,
    "steady_tpot_p95_ms_q25": 43.938638499999996,
    "steady_tpot_p95_ms_q75": 44.12715502499999,
    "decode_tokens_per_s": 20.805680993969492,
    "decode_tokens_per_s_q25": 20.381779865615734,
    "decode_tokens_per_s_q75": 20.918893179896052,
    "decode_ms": 720.9569350000002,
    "decode_ms_q25": 717.0761485000002,
    "decode_ms_q75": 736.2699080000002,
    "end_to_end_ms": 2427.292714,
    "end_to_end_ms_q25": 2409.8251550000004,
    "end_to_end_ms_q75": 2428.7121460000003,
    "output_tokens_per_s_e2e": 6.591706021987425,
    "output_tokens_per_s_e2e_q25": 6.587855827805798,
    "output_tokens_per_s_e2e_q75": 6.639834702265922,
    "all_tokens_per_s_e2e": 1687.4767416287807,
    "all_tokens_per_s_e2e_q25": 1686.4910919182844,
    "all_tokens_per_s_e2e_q75": 1699.797683780076,
    "energy_net_psys_j": 68.45919300439584,
    "energy_net_psys_j_q25": 68.34436544553681,
    "energy_net_psys_j_q75": 69.03062676639205,
    "prefill_energy_net_psys_j": 51.94530337567546,
    "prefill_energy_net_psys_j_q25": 51.755696855872586,
    "prefill_energy_net_psys_j_q75": 52.496566841306276,
    "decode_energy_net_psys_j": 16.55423022145115,
    "decode_energy_net_psys_j_q25": 16.41923236622673,
    "decode_energy_net_psys_j_q75": 16.723666444888632,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.963660938587566,
    "energy_reduction_vs_gpu": 0.05514876722070339
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 4096,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1647.3376349999999,
    "ttft_ms_q25": 1645.6815305,
    "ttft_ms_q75": 1649.6660885,
    "prefill_tokens_per_s": 2476.723601352069,
    "prefill_tokens_per_s_q25": 2473.2327086212017,
    "prefill_tokens_per_s_q75": 2479.2185220732376,
    "handoff_plus_first_decode_ms": 116.282667,
    "handoff_plus_first_decode_ms_q25": 113.9857415,
    "handoff_plus_first_decode_ms_q75": 132.1389845,
    "steady_tpot_ms": 42.9531705,
    "steady_tpot_ms_q25": 42.88008375,
    "steady_tpot_ms_q75": 43.00171325,
    "steady_tpot_p95_ms": 44.112806049999996,
    "steady_tpot_p95_ms_q25": 43.938638499999996,
    "steady_tpot_p95_ms_q75": 44.12715502499999,
    "decode_tokens_per_s": 20.805680993969492,
    "decode_tokens_per_s_q25": 20.381779865615734,
    "decode_tokens_per_s_q75": 20.918893179896052,
    "decode_ms": 720.9569350000002,
    "decode_ms_q25": 717.0761485000002,
    "decode_ms_q75": 736.2699080000002,
    "end_to_end_ms": 2372.951477,
    "end_to_end_ms_q25": 2365.0861325,
    "end_to_end_ms_q75": 2385.9359965000003,
    "output_tokens_per_s_e2e": 6.742657890429338,
    "output_tokens_per_s_e2e_q25": 6.7061622365399405,
    "output_tokens_per_s_e2e_q75": 6.765156133879357,
    "all_tokens_per_s_e2e": 1726.1204199499105,
    "all_tokens_per_s_e2e_q25": 1716.7775325542248,
    "all_tokens_per_s_e2e_q75": 1731.8799702731153,
    "energy_net_psys_j": 56.32058538076437,
    "energy_net_psys_j_q25": 56.17634495377227,
    "energy_net_psys_j_q75": 56.45572909818361,
    "prefill_energy_net_psys_j": 39.747870015777856,
    "prefill_energy_net_psys_j_q25": 39.7228200815273,
    "prefill_energy_net_psys_j_q75": 39.757112587545535,
    "decode_energy_net_psys_j": 16.55423022145115,
    "decode_energy_net_psys_j_q25": 16.41923236622673,
    "decode_energy_net_psys_j_q75": 16.723666444888632,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.9857290373072388,
    "energy_reduction_vs_gpu": 0.2226818314313188
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 4096,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1647.3376349999999,
    "ttft_ms_q25": 1645.6815305,
    "ttft_ms_q75": 1649.6660885,
    "prefill_tokens_per_s": 2476.723601352069,
    "prefill_tokens_per_s_q25": 2473.2327086212017,
    "prefill_tokens_per_s_q75": 2479.2185220732376,
    "handoff_plus_first_decode_ms": 60.345018,
    "handoff_plus_first_decode_ms_q25": 60.0690785,
    "handoff_plus_first_decode_ms_q75": 63.59227800000001,
    "steady_tpot_ms": 39.9686285,
    "steady_tpot_ms_q25": 39.96859825,
    "steady_tpot_ms_q75": 40.370734,
    "steady_tpot_p95_ms": 43.72544909999999,
    "steady_tpot_p95_ms_q25": 43.60656337499999,
    "steady_tpot_p95_ms_q75": 43.8626747,
    "decode_tokens_per_s": 23.812829937474554,
    "decode_tokens_per_s_q25": 23.522170856323314,
    "decode_tokens_per_s_q75": 23.832966702447706,
    "decode_ms": 629.9125320000005,
    "decode_ms_q25": 629.3807605000004,
    "decode_ms_q75": 637.7936290000004,
    "end_to_end_ms": 2281.9070740000006,
    "end_to_end_ms_q25": 2277.3907445000004,
    "end_to_end_ms_q75": 2287.4597175000004,
    "output_tokens_per_s_e2e": 7.011679039126374,
    "output_tokens_per_s_e2e_q25": 6.994699909531228,
    "output_tokens_per_s_e2e_q75": 7.025611640364678,
    "all_tokens_per_s_e2e": 1794.9898340163518,
    "all_tokens_per_s_e2e_q25": 1790.6431768399943,
    "all_tokens_per_s_e2e_q75": 1798.5565799333576,
    "energy_net_psys_j": 59.92741072206862,
    "energy_net_psys_j_q25": 59.67177025899524,
    "energy_net_psys_j_q75": 60.2570436438278,
    "prefill_energy_net_psys_j": 39.747870015777856,
    "prefill_energy_net_psys_j_q25": 39.7228200815273,
    "prefill_energy_net_psys_j_q75": 39.757112587545535,
    "decode_energy_net_psys_j": 20.1610555627554,
    "decode_energy_net_psys_j_q25": 19.914657671449696,
    "decode_energy_net_psys_j_q75": 20.52498099053282,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0250580322273017,
    "energy_reduction_vs_gpu": 0.1729016160856296
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 8192,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 4475.55968,
    "ttft_ms_q25": 4385.711507,
    "ttft_ms_q75": 4483.825789,
    "prefill_tokens_per_s": 1826.810630307582,
    "prefill_tokens_per_s_q25": 1823.4490303814532,
    "prefill_tokens_per_s_q75": 1865.0184560200641,
    "handoff_plus_first_decode_ms": 75.730897,
    "handoff_plus_first_decode_ms_q25": 74.952999,
    "handoff_plus_first_decode_ms_q75": 76.4329355,
    "steady_tpot_ms": 44.182173000000006,
    "steady_tpot_ms_q25": 44.0600525,
    "steady_tpot_ms_q75": 44.4365195,
    "steady_tpot_p95_ms": 47.9535477,
    "steady_tpot_p95_ms_q25": 47.95155835,
    "steady_tpot_p95_ms_q75": 48.401415575,
    "decode_tokens_per_s": 21.236507844865102,
    "decode_tokens_per_s_q25": 21.16574016738894,
    "decode_tokens_per_s_q75": 21.35928958761327,
    "decode_ms": 706.3308199999992,
    "decode_ms_q25": 702.2937544999995,
    "decode_ms_q75": 708.7003605,
    "end_to_end_ms": 5181.8904999999995,
    "end_to_end_ms_q25": 5094.4118675,
    "end_to_end_ms_q75": 5186.119543499999,
    "output_tokens_per_s_e2e": 3.087676206203122,
    "output_tokens_per_s_e2e_q25": 3.085160398878995,
    "output_tokens_per_s_e2e_q75": 3.141622540105343,
    "all_tokens_per_s_e2e": 1580.8902175759986,
    "all_tokens_per_s_e2e_q25": 1579.6021242260454,
    "all_tokens_per_s_e2e_q75": 1608.5107405339356,
    "energy_net_psys_j": 156.69767674481108,
    "energy_net_psys_j_q25": 154.8833029194545,
    "energy_net_psys_j_q75": 158.35906682980513,
    "prefill_energy_net_psys_j": 134.60066570578104,
    "prefill_energy_net_psys_j_q25": 132.68328291383347,
    "prefill_energy_net_psys_j_q75": 135.82319718374322,
    "decode_energy_net_psys_j": 22.303028972212015,
    "decode_energy_net_psys_j_q25": 22.20002000562102,
    "decode_energy_net_psys_j_q75": 22.638878612652892,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 8192,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 4475.55968,
    "ttft_ms_q25": 4385.711507,
    "ttft_ms_q75": 4483.825789,
    "prefill_tokens_per_s": 1826.810630307582,
    "prefill_tokens_per_s_q25": 1823.4490303814532,
    "prefill_tokens_per_s_q75": 1865.0184560200641,
    "handoff_plus_first_decode_ms": 152.325229,
    "handoff_plus_first_decode_ms_q25": 150.367657,
    "handoff_plus_first_decode_ms_q75": 161.8536205,
    "steady_tpot_ms": 47.188070499999995,
    "steady_tpot_ms_q25": 47.12804575,
    "steady_tpot_ms_q75": 47.229930249999995,
    "steady_tpot_p95_ms": 50.53907135000001,
    "steady_tpot_p95_ms_q25": 49.229488200000006,
    "steady_tpot_p95_ms_q75": 50.717617125000004,
    "decode_tokens_per_s": 18.334684491609448,
    "decode_tokens_per_s_q25": 18.093337313414104,
    "decode_tokens_per_s_q75": 18.369327745210455,
    "decode_ms": 818.1215230000007,
    "decode_ms_q25": 816.5815080000002,
    "decode_ms_q75": 829.1819885000004,
    "end_to_end_ms": 5293.681203000001,
    "end_to_end_ms_q25": 5214.8934955,
    "end_to_end_ms_q75": 5300.407297,
    "output_tokens_per_s_e2e": 3.022471393051131,
    "output_tokens_per_s_e2e_q25": 3.0186408078360607,
    "output_tokens_per_s_e2e_q75": 3.0688360128374303,
    "all_tokens_per_s_e2e": 1547.5053532421791,
    "all_tokens_per_s_e2e_q25": 1545.544093612063,
    "all_tokens_per_s_e2e_q75": 1571.2440385727643,
    "energy_net_psys_j": 152.65689247338162,
    "energy_net_psys_j_q25": 151.03686912498998,
    "energy_net_psys_j_q75": 153.24283607282825,
    "prefill_energy_net_psys_j": 134.60066570578104,
    "prefill_energy_net_psys_j_q25": 132.68328291383347,
    "prefill_energy_net_psys_j_q75": 135.82319718374322,
    "decode_energy_net_psys_j": 18.056226767600588,
    "decode_energy_net_psys_j_q25": 17.41963888908503,
    "decode_energy_net_psys_j_q75": 18.353586211156504,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9788822373858388,
    "energy_reduction_vs_gpu": 0.025787135810635187
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 8192,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 5123.838341,
    "ttft_ms_q25": 5118.5012145,
    "ttft_ms_q75": 5138.648222,
    "prefill_tokens_per_s": 1595.6787579688396,
    "prefill_tokens_per_s_q25": 1591.0931356454546,
    "prefill_tokens_per_s_q75": 1597.3443292742138,
    "handoff_plus_first_decode_ms": 152.325229,
    "handoff_plus_first_decode_ms_q25": 150.367657,
    "handoff_plus_first_decode_ms_q75": 161.8536205,
    "steady_tpot_ms": 47.188070499999995,
    "steady_tpot_ms_q25": 47.12804575,
    "steady_tpot_ms_q75": 47.229930249999995,
    "steady_tpot_p95_ms": 50.53907135000001,
    "steady_tpot_p95_ms_q25": 49.229488200000006,
    "steady_tpot_p95_ms_q75": 50.717617125000004,
    "decode_tokens_per_s": 18.334684491609448,
    "decode_tokens_per_s_q25": 18.093337313414104,
    "decode_tokens_per_s_q75": 18.369327745210455,
    "decode_ms": 818.1215230000007,
    "decode_ms_q25": 816.5815080000002,
    "decode_ms_q75": 829.1819885000004,
    "end_to_end_ms": 5953.406542000001,
    "end_to_end_ms_q25": 5946.143188,
    "end_to_end_ms_q75": 5962.493084000001,
    "output_tokens_per_s_e2e": 2.687536939922286,
    "output_tokens_per_s_e2e_q25": 2.6834474998058333,
    "output_tokens_per_s_e2e_q75": 2.6908238446464696,
    "all_tokens_per_s_e2e": 1376.0189132402104,
    "all_tokens_per_s_e2e_q25": 1373.9251199005867,
    "all_tokens_per_s_e2e_q75": 1377.7018084589924,
    "energy_net_psys_j": 124.69915985880303,
    "energy_net_psys_j_q25": 121.73285955947998,
    "energy_net_psys_j_q75": 125.55332830227859,
    "prefill_energy_net_psys_j": 106.04821420409061,
    "prefill_energy_net_psys_j_q25": 103.37927334832348,
    "prefill_energy_net_psys_j_q75": 107.83632996963765,
    "decode_energy_net_psys_j": 18.056226767600588,
    "decode_energy_net_psys_j_q25": 17.41963888908503,
    "decode_energy_net_psys_j_q75": 18.353586211156504,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.8704076335863977,
    "energy_reduction_vs_gpu": 0.20420543272073532
  },
  {
    "precision": "FP16",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 8192,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 5123.838341,
    "ttft_ms_q25": 5118.5012145,
    "ttft_ms_q75": 5138.648222,
    "prefill_tokens_per_s": 1595.6787579688396,
    "prefill_tokens_per_s_q25": 1591.0931356454546,
    "prefill_tokens_per_s_q75": 1597.3443292742138,
    "handoff_plus_first_decode_ms": 75.730897,
    "handoff_plus_first_decode_ms_q25": 74.952999,
    "handoff_plus_first_decode_ms_q75": 76.4329355,
    "steady_tpot_ms": 44.182173000000006,
    "steady_tpot_ms_q25": 44.0600525,
    "steady_tpot_ms_q75": 44.4365195,
    "steady_tpot_p95_ms": 47.9535477,
    "steady_tpot_p95_ms_q25": 47.95155835,
    "steady_tpot_p95_ms_q75": 48.401415575,
    "decode_tokens_per_s": 21.236507844865102,
    "decode_tokens_per_s_q25": 21.16574016738894,
    "decode_tokens_per_s_q75": 21.35928958761327,
    "decode_ms": 706.3308199999992,
    "decode_ms_q25": 702.2937544999995,
    "decode_ms_q75": 708.7003605,
    "end_to_end_ms": 5824.233989000001,
    "end_to_end_ms_q25": 5823.1645095,
    "end_to_end_ms_q75": 5842.011456,
    "output_tokens_per_s_e2e": 2.747142376185188,
    "output_tokens_per_s_e2e_q25": 2.738808077189325,
    "output_tokens_per_s_e2e_q75": 2.747647007669177,
    "all_tokens_per_s_e2e": 1406.5368966068163,
    "all_tokens_per_s_e2e_q25": 1402.2697355209343,
    "all_tokens_per_s_e2e_q75": 1406.7952679266186,
    "energy_net_psys_j": 128.35124317630263,
    "energy_net_psys_j_q25": 125.57929335394451,
    "energy_net_psys_j_q75": 130.47520858229052,
    "prefill_energy_net_psys_j": 106.04821420409061,
    "prefill_energy_net_psys_j_q25": 103.37927334832348,
    "prefill_energy_net_psys_j_q75": 107.83632996963765,
    "decode_energy_net_psys_j": 22.303028972212015,
    "decode_energy_net_psys_j_q25": 22.20002000562102,
    "decode_energy_net_psys_j_q75": 22.638878612652892,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.8897119363313407,
    "energy_reduction_vs_gpu": 0.1808988758312724
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 608.015883,
    "ttft_ms_q25": 592.866033,
    "ttft_ms_q75": 612.9221520000001,
    "prefill_tokens_per_s": 1657.851428200273,
    "prefill_tokens_per_s_q25": 1644.6861790917592,
    "prefill_tokens_per_s_q75": 1701.3264098583854,
    "handoff_plus_first_decode_ms": 96.843889,
    "handoff_plus_first_decode_ms_q25": 95.145674,
    "handoff_plus_first_decode_ms_q75": 101.84947550000001,
    "steady_tpot_ms": 86.96682750000001,
    "steady_tpot_ms_q25": 86.93742850000001,
    "steady_tpot_ms_q75": 86.9841755,
    "steady_tpot_p95_ms": 89.6742428,
    "steady_tpot_p95_ms_q25": 89.63911555,
    "steady_tpot_p95_ms_q75": 89.813657,
    "decode_tokens_per_s": 11.351113110507946,
    "decode_tokens_per_s_q25": 11.304144194500878,
    "decode_tokens_per_s_q75": 11.367143440480916,
    "decode_ms": 1321.456306,
    "decode_ms_q25": 1319.595368,
    "decode_ms_q75": 1326.969889,
    "end_to_end_ms": 1925.750313,
    "end_to_end_ms_q25": 1917.974984,
    "end_to_end_ms_q75": 1932.51752,
    "output_tokens_per_s_e2e": 8.308449902352423,
    "output_tokens_per_s_e2e_q25": 8.279457253900368,
    "output_tokens_per_s_e2e_q75": 8.342268844420449,
    "all_tokens_per_s_e2e": 531.7407937505551,
    "all_tokens_per_s_e2e_q25": 529.8852642496236,
    "all_tokens_per_s_e2e_q75": 533.9052060429087,
    "energy_net_psys_j": 65.31244432195557,
    "energy_net_psys_j_q25": 64.86409664222063,
    "energy_net_psys_j_q75": 65.45618557989684,
    "prefill_energy_net_psys_j": 22.669455582541488,
    "prefill_energy_net_psys_j_q25": 22.383034636950725,
    "prefill_energy_net_psys_j_q75": 22.80752426359583,
    "decode_energy_net_psys_j": 42.654333893187946,
    "decode_energy_net_psys_j_q25": 42.20031363656608,
    "decode_energy_net_psys_j_q75": 42.93508226189177,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 608.015883,
    "ttft_ms_q25": 592.866033,
    "ttft_ms_q75": 612.9221520000001,
    "prefill_tokens_per_s": 1657.851428200273,
    "prefill_tokens_per_s_q25": 1644.6861790917592,
    "prefill_tokens_per_s_q75": 1701.3264098583854,
    "handoff_plus_first_decode_ms": 109.747179,
    "handoff_plus_first_decode_ms_q25": 108.5420605,
    "handoff_plus_first_decode_ms_q75": 160.8111525,
    "steady_tpot_ms": 96.990364,
    "steady_tpot_ms_q25": 96.97914374999999,
    "steady_tpot_ms_q75": 97.06703775,
    "steady_tpot_p95_ms": 97.5769649,
    "steady_tpot_p95_ms_q25": 97.54373365000001,
    "steady_tpot_p95_ms_q75": 97.886416525,
    "decode_tokens_per_s": 10.202238934793973,
    "decode_tokens_per_s_q25": 9.875892479033057,
    "decode_tokens_per_s_q75": 10.207971083856268,
    "decode_ms": 1470.265507,
    "decode_ms_q25": 1469.4403625,
    "decode_ms_q75": 1520.510402,
    "end_to_end_ms": 2088.0939280000002,
    "end_to_end_ms_q25": 2082.3625145,
    "end_to_end_ms_q75": 2118.282704,
    "output_tokens_per_s_e2e": 7.662490554400002,
    "output_tokens_per_s_e2e_q25": 7.554822748193472,
    "output_tokens_per_s_e2e_q75": 7.683638702371958,
    "all_tokens_per_s_e2e": 490.39939548160015,
    "all_tokens_per_s_e2e_q25": 483.5086558843822,
    "all_tokens_per_s_e2e_q75": 491.7528769518053,
    "energy_net_psys_j": 53.469603405206925,
    "energy_net_psys_j_q25": 53.27340872764955,
    "energy_net_psys_j_q75": 53.7773644469636,
    "prefill_energy_net_psys_j": 22.669455582541488,
    "prefill_energy_net_psys_j_q25": 22.383034636950725,
    "prefill_energy_net_psys_j_q75": 22.80752426359583,
    "decode_energy_net_psys_j": 31.139532544070104,
    "decode_energy_net_psys_j_q25": 30.773645505810393,
    "decode_energy_net_psys_j_q75": 31.256261128958528,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9222527239684593,
    "energy_reduction_vs_gpu": 0.1813259485186275
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 764.6387970000001,
    "ttft_ms_q25": 762.237527,
    "ttft_ms_q75": 766.6386135,
    "prefill_tokens_per_s": 1318.2694939817445,
    "prefill_tokens_per_s_q25": 1314.8396666629214,
    "prefill_tokens_per_s_q75": 1322.4355507586504,
    "handoff_plus_first_decode_ms": 109.747179,
    "handoff_plus_first_decode_ms_q25": 108.5420605,
    "handoff_plus_first_decode_ms_q75": 160.8111525,
    "steady_tpot_ms": 96.990364,
    "steady_tpot_ms_q25": 96.97914374999999,
    "steady_tpot_ms_q75": 97.06703775,
    "steady_tpot_p95_ms": 97.5769649,
    "steady_tpot_p95_ms_q25": 97.54373365000001,
    "steady_tpot_p95_ms_q75": 97.886416525,
    "decode_tokens_per_s": 10.202238934793973,
    "decode_tokens_per_s_q25": 9.875892479033057,
    "decode_tokens_per_s_q75": 10.207971083856268,
    "decode_ms": 1470.265507,
    "decode_ms_q25": 1469.4403625,
    "decode_ms_q75": 1520.510402,
    "end_to_end_ms": 2234.904304,
    "end_to_end_ms_q25": 2231.6778894999998,
    "end_to_end_ms_q75": 2287.1490155,
    "output_tokens_per_s_e2e": 7.159143222089387,
    "output_tokens_per_s_e2e_q25": 6.999261040689813,
    "output_tokens_per_s_e2e_q75": 7.1695084302882,
    "all_tokens_per_s_e2e": 458.1851662137208,
    "all_tokens_per_s_e2e_q25": 447.95270660414803,
    "all_tokens_per_s_e2e_q75": 458.8485395384448,
    "energy_net_psys_j": 52.55018231241183,
    "energy_net_psys_j_q25": 52.39392568172434,
    "energy_net_psys_j_q75": 52.64115238927822,
    "prefill_energy_net_psys_j": 21.592589922074502,
    "prefill_energy_net_psys_j_q25": 21.384891260319687,
    "prefill_energy_net_psys_j_q75": 21.71125025278034,
    "decode_energy_net_psys_j": 31.139532544070104,
    "decode_energy_net_psys_j_q25": 30.773645505810393,
    "decode_energy_net_psys_j_q75": 31.256261128958528,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.8616701437969041,
    "energy_reduction_vs_gpu": 0.19540322127024656
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 764.6387970000001,
    "ttft_ms_q25": 762.237527,
    "ttft_ms_q75": 766.6386135,
    "prefill_tokens_per_s": 1318.2694939817445,
    "prefill_tokens_per_s_q25": 1314.8396666629214,
    "prefill_tokens_per_s_q75": 1322.4355507586504,
    "handoff_plus_first_decode_ms": 96.843889,
    "handoff_plus_first_decode_ms_q25": 95.145674,
    "handoff_plus_first_decode_ms_q75": 101.84947550000001,
    "steady_tpot_ms": 86.96682750000001,
    "steady_tpot_ms_q25": 86.93742850000001,
    "steady_tpot_ms_q75": 86.9841755,
    "steady_tpot_p95_ms": 89.6742428,
    "steady_tpot_p95_ms_q25": 89.63911555,
    "steady_tpot_p95_ms_q75": 89.813657,
    "decode_tokens_per_s": 11.351113110507946,
    "decode_tokens_per_s_q25": 11.304144194500878,
    "decode_tokens_per_s_q75": 11.367143440480916,
    "decode_ms": 1321.456306,
    "decode_ms_q25": 1319.595368,
    "decode_ms_q75": 1326.969889,
    "end_to_end_ms": 2086.095103,
    "end_to_end_ms_q25": 2081.832895,
    "end_to_end_ms_q75": 2093.6085025,
    "output_tokens_per_s_e2e": 7.669832490853606,
    "output_tokens_per_s_e2e_q25": 7.642405945279544,
    "output_tokens_per_s_e2e_q75": 7.685567417037216,
    "all_tokens_per_s_e2e": 490.8692794146308,
    "all_tokens_per_s_e2e_q25": 489.1139804978908,
    "all_tokens_per_s_e2e_q75": 491.87631469038183,
    "energy_net_psys_j": 64.24692381526245,
    "energy_net_psys_j_q25": 63.911563889346425,
    "energy_net_psys_j_q75": 64.31997352221146,
    "prefill_energy_net_psys_j": 21.592589922074502,
    "prefill_energy_net_psys_j_q25": 21.384891260319687,
    "prefill_energy_net_psys_j_q75": 21.71125025278034,
    "decode_energy_net_psys_j": 42.654333893187946,
    "decode_energy_net_psys_j_q25": 42.20031363656608,
    "decode_energy_net_psys_j_q75": 42.93508226189177,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9231363949949313,
    "energy_reduction_vs_gpu": 0.016314203483805745
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1333.976724,
    "ttft_ms_q25": 1306.1093925,
    "ttft_ms_q75": 1335.895748,
    "prefill_tokens_per_s": 1523.2649591568136,
    "prefill_tokens_per_s_q25": 1521.0799167457399,
    "prefill_tokens_per_s_q75": 1556.4741069968932,
    "handoff_plus_first_decode_ms": 100.18574600000001,
    "handoff_plus_first_decode_ms_q25": 99.9964105,
    "handoff_plus_first_decode_ms_q75": 102.64697749999999,
    "steady_tpot_ms": 87.75281500000001,
    "steady_tpot_ms_q25": 87.71656675,
    "steady_tpot_ms_q75": 87.79231925,
    "steady_tpot_p95_ms": 90.78791505,
    "steady_tpot_p95_ms_q25": 90.260761025,
    "steady_tpot_p95_ms_q75": 90.8908207,
    "decode_tokens_per_s": 11.225396339072269,
    "decode_tokens_per_s_q25": 11.211182400383763,
    "decode_tokens_per_s_q75": 11.227255763351272,
    "decode_ms": 1336.2557139999997,
    "decode_ms_q25": 1336.034444,
    "decode_ms_q75": 1337.9520174999996,
    "end_to_end_ms": 2669.789898,
    "end_to_end_ms_q25": 2643.84014,
    "end_to_end_ms_q75": 2671.930192,
    "output_tokens_per_s_e2e": 5.99298095029349,
    "output_tokens_per_s_e2e_q25": 5.988184240095618,
    "output_tokens_per_s_e2e_q75": 6.052386189881833,
    "all_tokens_per_s_e2e": 767.1015616375668,
    "all_tokens_per_s_e2e_q25": 766.487582732239,
    "all_tokens_per_s_e2e_q75": 774.7054323048746,
    "energy_net_psys_j": 92.22111851553885,
    "energy_net_psys_j_q25": 91.60348249114398,
    "energy_net_psys_j_q75": 92.42676630440494,
    "prefill_energy_net_psys_j": 49.39871092110475,
    "prefill_energy_net_psys_j_q25": 49.38592137439346,
    "prefill_energy_net_psys_j_q75": 49.69224627588552,
    "decode_energy_net_psys_j": 42.23533688487256,
    "decode_energy_net_psys_j_q25": 41.911236215258455,
    "decode_energy_net_psys_j_q75": 42.7473095752307,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1333.976724,
    "ttft_ms_q25": 1306.1093925,
    "ttft_ms_q75": 1335.895748,
    "prefill_tokens_per_s": 1523.2649591568136,
    "prefill_tokens_per_s_q25": 1521.0799167457399,
    "prefill_tokens_per_s_q75": 1556.4741069968932,
    "handoff_plus_first_decode_ms": 178.34850899999998,
    "handoff_plus_first_decode_ms_q25": 177.11055349999998,
    "handoff_plus_first_decode_ms_q75": 194.3620465,
    "steady_tpot_ms": 100.5010975,
    "steady_tpot_ms_q25": 100.4861975,
    "steady_tpot_ms_q75": 100.5453005,
    "steady_tpot_p95_ms": 101.3804163,
    "steady_tpot_p95_ms_q25": 101.075597725,
    "steady_tpot_p95_ms_q75": 101.993017025,
    "decode_tokens_per_s": 9.428149969795225,
    "decode_tokens_per_s_q25": 9.345168960724997,
    "decode_tokens_per_s_q75": 9.452399572268277,
    "decode_ms": 1590.980208,
    "decode_ms_q25": 1586.9090815,
    "decode_ms_q75": 1605.233983,
    "end_to_end_ms": 2920.652727,
    "end_to_end_ms_q25": 2909.1912730000004,
    "end_to_end_ms_q75": 2922.8048295,
    "output_tokens_per_s_e2e": 5.478227470211661,
    "output_tokens_per_s_e2e_q25": 5.474196741791277,
    "output_tokens_per_s_e2e_q75": 5.499895622894383,
    "all_tokens_per_s_e2e": 701.2131161870926,
    "all_tokens_per_s_e2e_q25": 700.6971829492835,
    "all_tokens_per_s_e2e_q75": 703.986639730481,
    "energy_net_psys_j": 81.06615447658498,
    "energy_net_psys_j_q25": 80.8737375662419,
    "energy_net_psys_j_q75": 81.11999194614235,
    "prefill_energy_net_psys_j": 49.39871092110475,
    "prefill_energy_net_psys_j_q25": 49.38592137439346,
    "prefill_energy_net_psys_j_q75": 49.69224627588552,
    "decode_energy_net_psys_j": 31.308188828216643,
    "decode_energy_net_psys_j_q25": 31.194280837067666,
    "decode_energy_net_psys_j_q75": 31.5416536614058,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9141072724323243,
    "energy_reduction_vs_gpu": 0.1209588890105936
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1576.823208,
    "ttft_ms_q25": 1575.7440855,
    "ttft_ms_q75": 1580.0888515000001,
    "prefill_tokens_per_s": 1288.6669790821597,
    "prefill_tokens_per_s_q25": 1286.0091239036803,
    "prefill_tokens_per_s_q75": 1289.550106328152,
    "handoff_plus_first_decode_ms": 178.34850899999998,
    "handoff_plus_first_decode_ms_q25": 177.11055349999998,
    "handoff_plus_first_decode_ms_q75": 194.3620465,
    "steady_tpot_ms": 100.5010975,
    "steady_tpot_ms_q25": 100.4861975,
    "steady_tpot_ms_q75": 100.5453005,
    "steady_tpot_p95_ms": 101.3804163,
    "steady_tpot_p95_ms_q25": 101.075597725,
    "steady_tpot_p95_ms_q75": 101.993017025,
    "decode_tokens_per_s": 9.428149969795225,
    "decode_tokens_per_s_q25": 9.345168960724997,
    "decode_tokens_per_s_q75": 9.452399572268277,
    "decode_ms": 1590.980208,
    "decode_ms_q25": 1586.9090815,
    "decode_ms_q75": 1605.233983,
    "end_to_end_ms": 3174.334703,
    "end_to_end_ms_q25": 3166.997933,
    "end_to_end_ms_q75": 3184.243712,
    "output_tokens_per_s_e2e": 5.040426261565525,
    "output_tokens_per_s_e2e_q25": 5.024789680114262,
    "output_tokens_per_s_e2e_q75": 5.052130190017129,
    "all_tokens_per_s_e2e": 645.1745614803872,
    "all_tokens_per_s_e2e_q25": 643.1730790546255,
    "all_tokens_per_s_e2e_q75": 646.6726643221925,
    "energy_net_psys_j": 76.77005017567933,
    "energy_net_psys_j_q25": 76.40717616521292,
    "energy_net_psys_j_q75": 77.37635485110076,
    "prefill_energy_net_psys_j": 45.46186134746268,
    "prefill_energy_net_psys_j_q25": 45.21289532814525,
    "prefill_energy_net_psys_j_q75": 45.834701189694954,
    "decode_energy_net_psys_j": 31.308188828216643,
    "decode_energy_net_psys_j_q25": 31.194280837067666,
    "decode_energy_net_psys_j_q75": 31.5416536614058,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.8410549446713464,
    "energy_reduction_vs_gpu": 0.16754371003704627
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1576.823208,
    "ttft_ms_q25": 1575.7440855,
    "ttft_ms_q75": 1580.0888515000001,
    "prefill_tokens_per_s": 1288.6669790821597,
    "prefill_tokens_per_s_q25": 1286.0091239036803,
    "prefill_tokens_per_s_q75": 1289.550106328152,
    "handoff_plus_first_decode_ms": 100.18574600000001,
    "handoff_plus_first_decode_ms_q25": 99.9964105,
    "handoff_plus_first_decode_ms_q75": 102.64697749999999,
    "steady_tpot_ms": 87.75281500000001,
    "steady_tpot_ms_q25": 87.71656675,
    "steady_tpot_ms_q75": 87.79231925,
    "steady_tpot_p95_ms": 90.78791505,
    "steady_tpot_p95_ms_q25": 90.260761025,
    "steady_tpot_p95_ms_q75": 90.8908207,
    "decode_tokens_per_s": 11.225396339072269,
    "decode_tokens_per_s_q25": 11.211182400383763,
    "decode_tokens_per_s_q75": 11.227255763351272,
    "decode_ms": 1336.2557139999997,
    "decode_ms_q25": 1336.034444,
    "decode_ms_q75": 1337.9520174999996,
    "end_to_end_ms": 2914.3132839999994,
    "end_to_end_ms_q25": 2913.6961029999993,
    "end_to_end_ms_q75": 2916.7404764999997,
    "output_tokens_per_s_e2e": 5.490144140591305,
    "output_tokens_per_s_e2e_q25": 5.48557926519087,
    "output_tokens_per_s_e2e_q75": 5.4913073128710685,
    "all_tokens_per_s_e2e": 702.738449995687,
    "all_tokens_per_s_e2e_q25": 702.1541459444313,
    "all_tokens_per_s_e2e_q75": 702.8873360474968,
    "energy_net_psys_j": 87.7946765775716,
    "energy_net_psys_j_q25": 87.49697138563599,
    "energy_net_psys_j_q75": 88.25791009531156,
    "prefill_energy_net_psys_j": 45.46186134746268,
    "prefill_energy_net_psys_j_q25": 45.21289532814525,
    "prefill_energy_net_psys_j_q75": 45.834701189694954,
    "decode_energy_net_psys_j": 42.23533688487256,
    "decode_energy_net_psys_j_q25": 41.911236215258455,
    "decode_energy_net_psys_j_q75": 42.7473095752307,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9160957103196599,
    "energy_reduction_vs_gpu": 0.04799813761987082
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3401.502968,
    "ttft_ms_q25": 3349.8570529999997,
    "ttft_ms_q75": 3403.8467545,
    "prefill_tokens_per_s": 1199.4697750914913,
    "prefill_tokens_per_s_q25": 1198.6444245915184,
    "prefill_tokens_per_s_q75": 1218.2519911042018,
    "handoff_plus_first_decode_ms": 109.114073,
    "handoff_plus_first_decode_ms_q25": 108.62187650000001,
    "handoff_plus_first_decode_ms_q75": 111.657108,
    "steady_tpot_ms": 92.99677700000001,
    "steady_tpot_ms_q25": 92.99468750000001,
    "steady_tpot_ms_q75": 93.00464825,
    "steady_tpot_p95_ms": 94.43623765000001,
    "steady_tpot_p95_ms_q25": 94.29474630000001,
    "steady_tpot_p95_ms_q75": 94.71633805,
    "decode_tokens_per_s": 10.594658249592095,
    "decode_tokens_per_s_q25": 10.580235900469772,
    "decode_tokens_per_s_q75": 10.604553138881375,
    "decode_ms": 1415.807820000001,
    "decode_ms_q25": 1414.4879905000003,
    "decode_ms_q75": 1417.7403995000006,
    "end_to_end_ms": 4817.310788000001,
    "end_to_end_ms_q25": 4767.5974525,
    "end_to_end_ms_q75": 4818.334745,
    "output_tokens_per_s_e2e": 3.3213551510640045,
    "output_tokens_per_s_e2e_q25": 3.3206494711039287,
    "output_tokens_per_s_e2e_q75": 3.356352967637018,
    "all_tokens_per_s_e2e": 850.2669186723851,
    "all_tokens_per_s_e2e_q25": 850.0862646026058,
    "all_tokens_per_s_e2e_q75": 859.2263597150766,
    "energy_net_psys_j": 165.47752763595605,
    "energy_net_psys_j_q25": 162.6926125645806,
    "energy_net_psys_j_q75": 166.65769336677786,
    "prefill_energy_net_psys_j": 120.24855199467956,
    "prefill_energy_net_psys_j_q25": 118.71971973437182,
    "prefill_energy_net_psys_j_q75": 121.24681045981217,
    "decode_energy_net_psys_j": 45.22897564127649,
    "decode_energy_net_psys_j_q25": 43.97289283020879,
    "decode_energy_net_psys_j_q75": 45.41088290696568,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3401.502968,
    "ttft_ms_q25": 3349.8570529999997,
    "ttft_ms_q75": 3403.8467545,
    "prefill_tokens_per_s": 1199.4697750914913,
    "prefill_tokens_per_s_q25": 1198.6444245915184,
    "prefill_tokens_per_s_q75": 1218.2519911042018,
    "handoff_plus_first_decode_ms": 266.97515,
    "handoff_plus_first_decode_ms_q25": 263.0479305,
    "handoff_plus_first_decode_ms_q75": 275.74849299999994,
    "steady_tpot_ms": 107.17812599999999,
    "steady_tpot_ms_q25": 107.17021025,
    "steady_tpot_ms_q75": 107.49997124999999,
    "steady_tpot_p95_ms": 109.60956010000001,
    "steady_tpot_p95_ms_q25": 109.233686425,
    "steady_tpot_p95_ms_q75": 109.8088157,
    "decode_tokens_per_s": 8.466271671821122,
    "decode_tokens_per_s_q25": 8.401582366719186,
    "decode_tokens_per_s_q75": 8.480237867743881,
    "decode_ms": 1771.7361999999998,
    "decode_ms_q25": 1768.823106,
    "decode_ms_q75": 1785.4838134999998,
    "end_to_end_ms": 5167.41298,
    "end_to_end_ms_q25": 5132.427772499999,
    "end_to_end_ms_q75": 5172.6698605,
    "output_tokens_per_s_e2e": 3.0963269361141714,
    "output_tokens_per_s_e2e_q25": 3.093183395947178,
    "output_tokens_per_s_e2e_q75": 3.1175779142206395,
    "all_tokens_per_s_e2e": 792.6596956452279,
    "all_tokens_per_s_e2e_q25": 791.8549493624776,
    "all_tokens_per_s_e2e_q75": 798.0999460404837,
    "energy_net_psys_j": 153.62109845076316,
    "energy_net_psys_j_q25": 152.04089340171447,
    "energy_net_psys_j_q75": 155.1668899905201,
    "prefill_energy_net_psys_j": 120.24855199467956,
    "prefill_energy_net_psys_j_q25": 118.71971973437182,
    "prefill_energy_net_psys_j_q75": 121.24681045981217,
    "decode_energy_net_psys_j": 33.3725464560836,
    "decode_energy_net_psys_j_q25": 33.321173667342634,
    "decode_energy_net_psys_j_q75": 33.92007953070794,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9322480720323617,
    "energy_reduction_vs_gpu": 0.07164978444249281
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3819.786142,
    "ttft_ms_q25": 3813.6422555,
    "ttft_ms_q75": 3821.775001,
    "prefill_tokens_per_s": 1068.1226247560955,
    "prefill_tokens_per_s_q25": 1067.5670608214152,
    "prefill_tokens_per_s_q75": 1069.8461774772245,
    "handoff_plus_first_decode_ms": 266.97515,
    "handoff_plus_first_decode_ms_q25": 263.0479305,
    "handoff_plus_first_decode_ms_q75": 275.74849299999994,
    "steady_tpot_ms": 107.17812599999999,
    "steady_tpot_ms_q25": 107.17021025,
    "steady_tpot_ms_q75": 107.49997124999999,
    "steady_tpot_p95_ms": 109.60956010000001,
    "steady_tpot_p95_ms_q25": 109.233686425,
    "steady_tpot_p95_ms_q75": 109.8088157,
    "decode_tokens_per_s": 8.466271671821122,
    "decode_tokens_per_s_q25": 8.401582366719186,
    "decode_tokens_per_s_q75": 8.480237867743881,
    "decode_ms": 1771.7361999999998,
    "decode_ms_q25": 1768.823106,
    "decode_ms_q75": 1785.4838134999998,
    "end_to_end_ms": 5589.673872,
    "end_to_end_ms_q25": 5584.4542205,
    "end_to_end_ms_q75": 5604.3457204999995,
    "output_tokens_per_s_e2e": 2.862421022476425,
    "output_tokens_per_s_e2e_q25": 2.8549469387830966,
    "output_tokens_per_s_e2e_q75": 2.8650989597452066,
    "all_tokens_per_s_e2e": 732.7797817539648,
    "all_tokens_per_s_e2e_q25": 730.8664163284727,
    "all_tokens_per_s_e2e_q75": 733.4653336947729,
    "energy_net_psys_j": 138.65458698340248,
    "energy_net_psys_j_q25": 138.23103475952297,
    "energy_net_psys_j_q75": 140.5430702314295,
    "prefill_energy_net_psys_j": 105.28204052731888,
    "prefill_energy_net_psys_j_q25": 104.90986109218036,
    "prefill_energy_net_psys_j_q75": 106.62299070072156,
    "decode_energy_net_psys_j": 33.3725464560836,
    "decode_energy_net_psys_j_q25": 33.321173667342634,
    "decode_energy_net_psys_j_q75": 33.92007953070794,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.8618232294608548,
    "energy_reduction_vs_gpu": 0.16209415886103262
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3819.786142,
    "ttft_ms_q25": 3813.6422555,
    "ttft_ms_q75": 3821.775001,
    "prefill_tokens_per_s": 1068.1226247560955,
    "prefill_tokens_per_s_q25": 1067.5670608214152,
    "prefill_tokens_per_s_q75": 1069.8461774772245,
    "handoff_plus_first_decode_ms": 109.114073,
    "handoff_plus_first_decode_ms_q25": 108.62187650000001,
    "handoff_plus_first_decode_ms_q75": 111.657108,
    "steady_tpot_ms": 92.99677700000001,
    "steady_tpot_ms_q25": 92.99468750000001,
    "steady_tpot_ms_q75": 93.00464825,
    "steady_tpot_p95_ms": 94.43623765000001,
    "steady_tpot_p95_ms_q25": 94.29474630000001,
    "steady_tpot_p95_ms_q75": 94.71633805,
    "decode_tokens_per_s": 10.594658249592095,
    "decode_tokens_per_s_q25": 10.580235900469772,
    "decode_tokens_per_s_q75": 10.604553138881375,
    "decode_ms": 1415.807820000001,
    "decode_ms_q25": 1414.4879905000003,
    "decode_ms_q75": 1417.7403995000006,
    "end_to_end_ms": 5239.459121,
    "end_to_end_ms_q25": 5230.0628255,
    "end_to_end_ms_q75": 5239.5154005,
    "output_tokens_per_s_e2e": 3.053750326225706,
    "output_tokens_per_s_e2e_q25": 3.053717525158752,
    "output_tokens_per_s_e2e_q75": 3.059246547883142,
    "all_tokens_per_s_e2e": 781.7600835137807,
    "all_tokens_per_s_e2e_q25": 781.7516864406405,
    "all_tokens_per_s_e2e_q75": 783.1671162580843,
    "energy_net_psys_j": 150.51101616859538,
    "energy_net_psys_j_q25": 148.88275392238916,
    "energy_net_psys_j_q75": 152.03387360768724,
    "prefill_energy_net_psys_j": 105.28204052731888,
    "prefill_energy_net_psys_j_q25": 104.90986109218036,
    "prefill_energy_net_psys_j_q75": 106.62299070072156,
    "decode_energy_net_psys_j": 45.22897564127649,
    "decode_energy_net_psys_j_q25": 43.97289283020879,
    "decode_energy_net_psys_j_q75": 45.41088290696568,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9194290243991009,
    "energy_reduction_vs_gpu": 0.09044437441853981
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 8393.731064000001,
    "ttft_ms_q25": 8304.008112,
    "ttft_ms_q75": 8398.885063000002,
    "prefill_tokens_per_s": 974.0602763729432,
    "prefill_tokens_per_s_q25": 973.4629081725661,
    "prefill_tokens_per_s_q75": 984.699736985742,
    "handoff_plus_first_decode_ms": 124.933821,
    "handoff_plus_first_decode_ms_q25": 124.2429285,
    "handoff_plus_first_decode_ms_q75": 126.7821945,
    "steady_tpot_ms": 99.4496485,
    "steady_tpot_ms_q25": 99.42917675,
    "steady_tpot_ms_q75": 99.571631,
    "steady_tpot_p95_ms": 101.72606005,
    "steady_tpot_p95_ms_q25": 101.66135237499999,
    "steady_tpot_p95_ms_q75": 101.957472725,
    "decode_tokens_per_s": 9.825767383491202,
    "decode_tokens_per_s_q25": 9.823172126265263,
    "decode_tokens_per_s_q75": 9.843393928516315,
    "decode_ms": 1526.5983220000016,
    "decode_ms_q25": 1523.8695320000006,
    "decode_ms_q75": 1527.0017520000001,
    "end_to_end_ms": 9914.871806000001,
    "end_to_end_ms_q25": 9828.281073999999,
    "end_to_end_ms_q75": 9922.754595000002,
    "output_tokens_per_s_e2e": 1.6137374555178388,
    "output_tokens_per_s_e2e_q25": 1.6124564952560392,
    "output_tokens_per_s_e2e_q75": 1.6280814453823762,
    "all_tokens_per_s_e2e": 826.2335772251334,
    "all_tokens_per_s_e2e_q25": 825.5777255710921,
    "all_tokens_per_s_e2e_q75": 833.5777000357766,
    "energy_net_psys_j": 343.93872694202236,
    "energy_net_psys_j_q25": 342.82446973193373,
    "energy_net_psys_j_q75": 345.7816045764536,
    "prefill_energy_net_psys_j": 295.5721234930767,
    "prefill_energy_net_psys_j_q25": 294.5815669048591,
    "prefill_energy_net_psys_j_q75": 297.14477334743196,
    "decode_energy_net_psys_j": 48.36660344894568,
    "decode_energy_net_psys_j_q25": 48.242902827074644,
    "decode_energy_net_psys_j_q75": 48.63683122902165,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 8393.731064000001,
    "ttft_ms_q25": 8304.008112,
    "ttft_ms_q75": 8398.885063000002,
    "prefill_tokens_per_s": 974.0602763729432,
    "prefill_tokens_per_s_q25": 973.4629081725661,
    "prefill_tokens_per_s_q75": 984.699736985742,
    "handoff_plus_first_decode_ms": 396.897302,
    "handoff_plus_first_decode_ms_q25": 396.0727155,
    "handoff_plus_first_decode_ms_q75": 411.564028,
    "steady_tpot_ms": 123.84031549999999,
    "steady_tpot_ms_q25": 123.51343499999999,
    "steady_tpot_ms_q75": 123.859394,
    "steady_tpot_p95_ms": 125.85470939999999,
    "steady_tpot_p95_ms_q25": 125.586965825,
    "steady_tpot_p95_ms_q75": 126.02089537500001,
    "decode_tokens_per_s": 7.025409807563141,
    "decode_tokens_per_s_q25": 6.989961835045682,
    "decode_tokens_per_s_q75": 7.025609047548404,
    "decode_ms": 2135.106764,
    "decode_ms_q25": 2135.0462160000006,
    "decode_ms_q75": 2145.9896535000007,
    "end_to_end_ms": 10528.716732000003,
    "end_to_end_ms_q25": 10449.937217500003,
    "end_to_end_ms_q75": 10533.931279,
    "output_tokens_per_s_e2e": 1.519653382958921,
    "output_tokens_per_s_e2e_q25": 1.5189014905266522,
    "output_tokens_per_s_e2e_q75": 1.5311966994474127,
    "all_tokens_per_s_e2e": 778.0625320749675,
    "all_tokens_per_s_e2e_q25": 777.677563149646,
    "all_tokens_per_s_e2e_q75": 783.9727101170753,
    "energy_net_psys_j": 333.11812980832195,
    "energy_net_psys_j_q25": 332.83437756564086,
    "energy_net_psys_j_q75": 334.9759381692131,
    "prefill_energy_net_psys_j": 295.5721234930767,
    "prefill_energy_net_psys_j_q25": 294.5815669048591,
    "prefill_energy_net_psys_j_q75": 297.14477334743196,
    "decode_energy_net_psys_j": 38.116323328317065,
    "decode_energy_net_psys_j_q25": 37.83116482178118,
    "decode_energy_net_psys_j_q75": 38.53796916731765,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9416980300994956,
    "energy_reduction_vs_gpu": 0.031460828008252895
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 9824.381988000001,
    "ttft_ms_q25": 9823.482581,
    "ttft_ms_q75": 9832.498856,
    "prefill_tokens_per_s": 832.2151978604437,
    "prefill_tokens_per_s_q25": 831.5287590137207,
    "prefill_tokens_per_s_q75": 832.2913998290301,
    "handoff_plus_first_decode_ms": 396.897302,
    "handoff_plus_first_decode_ms_q25": 396.0727155,
    "handoff_plus_first_decode_ms_q75": 411.564028,
    "steady_tpot_ms": 123.84031549999999,
    "steady_tpot_ms_q25": 123.51343499999999,
    "steady_tpot_ms_q75": 123.859394,
    "steady_tpot_p95_ms": 125.85470939999999,
    "steady_tpot_p95_ms_q25": 125.586965825,
    "steady_tpot_p95_ms_q75": 126.02089537500001,
    "decode_tokens_per_s": 7.025409807563141,
    "decode_tokens_per_s_q25": 6.989961835045682,
    "decode_tokens_per_s_q75": 7.025609047548404,
    "decode_ms": 2135.106764,
    "decode_ms_q25": 2135.0462160000006,
    "decode_ms_q75": 2145.9896535000007,
    "end_to_end_ms": 11975.722488,
    "end_to_end_ms_q25": 11967.545072,
    "end_to_end_ms_q75": 11977.5891025,
    "output_tokens_per_s_e2e": 1.3360363031151095,
    "output_tokens_per_s_e2e_q25": 1.3358281246473362,
    "output_tokens_per_s_e2e_q75": 1.3369498400976598,
    "all_tokens_per_s_e2e": 684.050587194936,
    "all_tokens_per_s_e2e_q25": 683.9439998194362,
    "all_tokens_per_s_e2e_q75": 684.5183181300018,
    "energy_net_psys_j": 296.3685678572622,
    "energy_net_psys_j_q25": 294.1043334975039,
    "energy_net_psys_j_q75": 296.77729115229215,
    "prefill_energy_net_psys_j": 258.22639944100393,
    "prefill_energy_net_psys_j_q25": 256.2602461317521,
    "prefill_energy_net_psys_j_q75": 258.23932198497454,
    "decode_energy_net_psys_j": 38.116323328317065,
    "decode_energy_net_psys_j_q25": 37.83116482178118,
    "decode_energy_net_psys_j_q75": 38.53796916731765,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.8279142920967794,
    "energy_reduction_vs_gpu": 0.13830998186133037
  },
  {
    "precision": "FP16",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 9824.381988000001,
    "ttft_ms_q25": 9823.482581,
    "ttft_ms_q75": 9832.498856,
    "prefill_tokens_per_s": 832.2151978604437,
    "prefill_tokens_per_s_q25": 831.5287590137207,
    "prefill_tokens_per_s_q75": 832.2913998290301,
    "handoff_plus_first_decode_ms": 124.933821,
    "handoff_plus_first_decode_ms_q25": 124.2429285,
    "handoff_plus_first_decode_ms_q75": 126.7821945,
    "steady_tpot_ms": 99.4496485,
    "steady_tpot_ms_q25": 99.42917675,
    "steady_tpot_ms_q75": 99.571631,
    "steady_tpot_p95_ms": 101.72606005,
    "steady_tpot_p95_ms_q25": 101.66135237499999,
    "steady_tpot_p95_ms_q75": 101.957472725,
    "decode_tokens_per_s": 9.825767383491202,
    "decode_tokens_per_s_q25": 9.823172126265263,
    "decode_tokens_per_s_q75": 9.843393928516315,
    "decode_ms": 1526.5983220000016,
    "decode_ms_q25": 1523.8695320000006,
    "decode_ms_q75": 1527.0017520000001,
    "end_to_end_ms": 11349.988355999998,
    "end_to_end_ms_q25": 11347.755543,
    "end_to_end_ms_q75": 11358.601201,
    "output_tokens_per_s_e2e": 1.4096930761644213,
    "output_tokens_per_s_e2e_q25": 1.4086249632641588,
    "output_tokens_per_s_e2e_q75": 1.4099705055170721,
    "all_tokens_per_s_e2e": 721.7628549961837,
    "all_tokens_per_s_e2e_q25": 721.2159811912493,
    "all_tokens_per_s_e2e_q75": 721.9048988247409,
    "energy_net_psys_j": 306.34560164620757,
    "energy_net_psys_j_q25": 304.5031489588268,
    "energy_net_psys_j_q75": 306.75245259212517,
    "prefill_energy_net_psys_j": 258.22639944100393,
    "prefill_energy_net_psys_j_q25": 256.2602461317521,
    "prefill_energy_net_psys_j_q75": 258.23932198497454,
    "decode_energy_net_psys_j": 48.36660344894568,
    "decode_energy_net_psys_j_q25": 48.242902827074644,
    "decode_energy_net_psys_j_q75": 48.63683122902165,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.873557883498502,
    "energy_reduction_vs_gpu": 0.10930180974401249
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 1024,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 207.69891900000002,
    "ttft_ms_q25": 194.50003650000002,
    "ttft_ms_q75": 207.97992950000003,
    "prefill_tokens_per_s": 4853.1788458658275,
    "prefill_tokens_per_s_q25": 4846.6303581174,
    "prefill_tokens_per_s_q75": 5206.494508368829,
    "handoff_plus_first_decode_ms": 24.224101,
    "handoff_plus_first_decode_ms_q25": 23.6317355,
    "handoff_plus_first_decode_ms_q75": 30.771002000000003,
    "steady_tpot_ms": 15.2328625,
    "steady_tpot_ms_q25": 15.22344175,
    "steady_tpot_ms_q75": 15.295879750000001,
    "steady_tpot_p95_ms": 16.608904,
    "steady_tpot_p95_ms_q25": 16.084898199999998,
    "steady_tpot_p95_ms_q75": 17.133307324999997,
    "decode_tokens_per_s": 61.171514515309156,
    "decode_tokens_per_s_q25": 60.47460600914057,
    "decode_tokens_per_s_q75": 61.57064257445532,
    "decode_ms": 245.21217299999998,
    "decode_ms_q25": 243.6328375,
    "decode_ms_q75": 248.070939,
    "end_to_end_ms": 450.31444200000004,
    "end_to_end_ms_q25": 441.27265050000005,
    "end_to_end_ms_q75": 451.612767,
    "output_tokens_per_s_e2e": 35.53072810398561,
    "output_tokens_per_s_e2e_q25": 35.428874929310865,
    "output_tokens_per_s_e2e_q75": 36.27399163704915,
    "all_tokens_per_s_e2e": 2273.966598655079,
    "all_tokens_per_s_e2e_q25": 2267.4479954758954,
    "all_tokens_per_s_e2e_q75": 2321.5354647711456,
    "energy_net_psys_j": 14.367050427195586,
    "energy_net_psys_j_q25": 14.34006304929965,
    "energy_net_psys_j_q75": 14.397856528065999,
    "prefill_energy_net_psys_j": 5.535363733573653,
    "prefill_energy_net_psys_j_q25": 5.0970728583691605,
    "prefill_energy_net_psys_j_q75": 5.543861215840693,
    "decode_energy_net_psys_j": 8.876303930828678,
    "decode_energy_net_psys_j_q25": 8.853995312225305,
    "decode_energy_net_psys_j_q75": 9.265298809533862,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 1024,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 207.69891900000002,
    "ttft_ms_q25": 194.50003650000002,
    "ttft_ms_q75": 207.97992950000003,
    "prefill_tokens_per_s": 4853.1788458658275,
    "prefill_tokens_per_s_q25": 4846.6303581174,
    "prefill_tokens_per_s_q75": 5206.494508368829,
    "handoff_plus_first_decode_ms": 21.64219,
    "handoff_plus_first_decode_ms_q25": 20.996133,
    "handoff_plus_first_decode_ms_q75": 56.0067405,
    "steady_tpot_ms": 16.6017975,
    "steady_tpot_ms_q25": 16.534877,
    "steady_tpot_ms_q75": 16.622454750000003,
    "steady_tpot_p95_ms": 17.67764695,
    "steady_tpot_p95_ms_q25": 17.65987035,
    "steady_tpot_p95_ms_q75": 17.85497235,
    "decode_tokens_per_s": 58.259557701827305,
    "decode_tokens_per_s_q25": 52.105704847208315,
    "decode_tokens_per_s_q75": 58.40052970577568,
    "decode_ms": 257.468484,
    "decode_ms_q25": 256.848482,
    "decode_ms_q75": 291.94855600000005,
    "end_to_end_ms": 465.16740300000004,
    "end_to_end_ms_q25": 464.8284115,
    "end_to_end_ms_q75": 486.4485925,
    "output_tokens_per_s_e2e": 34.39621928968226,
    "output_tokens_per_s_e2e_q25": 32.95452237384665,
    "output_tokens_per_s_e2e_q75": 34.421322177862876,
    "all_tokens_per_s_e2e": 2201.3580345396645,
    "all_tokens_per_s_e2e_q25": 2109.0894319261856,
    "all_tokens_per_s_e2e_q75": 2202.964619383224,
    "energy_net_psys_j": 12.934269199598155,
    "energy_net_psys_j_q25": 12.890027974147833,
    "energy_net_psys_j_q75": 13.328539196585298,
    "prefill_energy_net_psys_j": 5.535363733573653,
    "prefill_energy_net_psys_j_q25": 5.0970728583691605,
    "prefill_energy_net_psys_j_q75": 5.543861215840693,
    "decode_energy_net_psys_j": 7.381910501490423,
    "decode_energy_net_psys_j_q25": 7.34616675830714,
    "decode_energy_net_psys_j_q75": 8.222968855949098,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9680696435214313,
    "energy_reduction_vs_gpu": 0.09972688791328388
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 1024,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 203.698701,
    "ttft_ms_q25": 202.494179,
    "ttft_ms_q75": 203.72010649999999,
    "prefill_tokens_per_s": 4948.485164861213,
    "prefill_tokens_per_s_q25": 4947.9652668877625,
    "prefill_tokens_per_s_q75": 4978.0970154479965,
    "handoff_plus_first_decode_ms": 21.64219,
    "handoff_plus_first_decode_ms_q25": 20.996133,
    "handoff_plus_first_decode_ms_q75": 56.0067405,
    "steady_tpot_ms": 16.6017975,
    "steady_tpot_ms_q25": 16.534877,
    "steady_tpot_ms_q75": 16.622454750000003,
    "steady_tpot_p95_ms": 17.67764695,
    "steady_tpot_p95_ms_q25": 17.65987035,
    "steady_tpot_p95_ms_q75": 17.85497235,
    "decode_tokens_per_s": 58.259557701827305,
    "decode_tokens_per_s_q25": 52.105704847208315,
    "decode_tokens_per_s_q75": 58.40052970577568,
    "decode_ms": 257.468484,
    "decode_ms_q25": 256.848482,
    "decode_ms_q75": 291.94855600000005,
    "end_to_end_ms": 461.167185,
    "end_to_end_ms_q25": 459.342661,
    "end_to_end_ms_q75": 495.66866250000004,
    "output_tokens_per_s_e2e": 34.69457611126429,
    "output_tokens_per_s_e2e_q25": 32.43678366547543,
    "output_tokens_per_s_e2e_q75": 34.83293365748071,
    "all_tokens_per_s_e2e": 2220.4528711209146,
    "all_tokens_per_s_e2e_q25": 2075.9541545904276,
    "all_tokens_per_s_e2e_q75": 2229.3077540787654,
    "energy_net_psys_j": 10.16254561750419,
    "energy_net_psys_j_q25": 10.086285060208787,
    "energy_net_psys_j_q75": 10.78111545520396,
    "prefill_energy_net_psys_j": 2.6281140014229605,
    "prefill_energy_net_psys_j_q25": 2.4818860419594584,
    "prefill_energy_net_psys_j_q75": 2.7401183019016466,
    "decode_energy_net_psys_j": 7.381910501490423,
    "decode_energy_net_psys_j_q25": 7.34616675830714,
    "decode_energy_net_psys_j_q75": 8.222968855949098,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.9764667926231568,
    "energy_reduction_vs_gpu": 0.2926491301048566
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 1024,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 203.698701,
    "ttft_ms_q25": 202.494179,
    "ttft_ms_q75": 203.72010649999999,
    "prefill_tokens_per_s": 4948.485164861213,
    "prefill_tokens_per_s_q25": 4947.9652668877625,
    "prefill_tokens_per_s_q75": 4978.0970154479965,
    "handoff_plus_first_decode_ms": 24.224101,
    "handoff_plus_first_decode_ms_q25": 23.6317355,
    "handoff_plus_first_decode_ms_q75": 30.771002000000003,
    "steady_tpot_ms": 15.2328625,
    "steady_tpot_ms_q25": 15.22344175,
    "steady_tpot_ms_q75": 15.295879750000001,
    "steady_tpot_p95_ms": 16.608904,
    "steady_tpot_p95_ms_q25": 16.084898199999998,
    "steady_tpot_p95_ms_q75": 17.133307324999997,
    "decode_tokens_per_s": 61.171514515309156,
    "decode_tokens_per_s_q25": 60.47460600914057,
    "decode_tokens_per_s_q75": 61.57064257445532,
    "decode_ms": 245.21217299999998,
    "decode_ms_q25": 243.6328375,
    "decode_ms_q75": 248.070939,
    "end_to_end_ms": 448.910874,
    "end_to_end_ms_q25": 446.12701649999997,
    "end_to_end_ms_q75": 451.7910455,
    "output_tokens_per_s_e2e": 35.6418187366074,
    "output_tokens_per_s_e2e_q25": 35.41604119297779,
    "output_tokens_per_s_e2e_q75": 35.86562222243843,
    "all_tokens_per_s_e2e": 2281.0763991428735,
    "all_tokens_per_s_e2e_q25": 2266.6266363505783,
    "all_tokens_per_s_e2e_q75": 2295.3998222360597,
    "energy_net_psys_j": 11.683809296002266,
    "energy_net_psys_j_q25": 11.594113614126952,
    "energy_net_psys_j_q75": 11.836880533368634,
    "prefill_energy_net_psys_j": 2.6281140014229605,
    "prefill_energy_net_psys_j_q25": 2.4818860419594584,
    "prefill_energy_net_psys_j_q75": 2.7401183019016466,
    "decode_energy_net_psys_j": 8.876303930828678,
    "decode_energy_net_psys_j_q25": 8.853995312225305,
    "decode_energy_net_psys_j_q75": 9.265298809533862,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0031266072650316,
    "energy_reduction_vs_gpu": 0.18676353540975787
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 2048,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 441.579209,
    "ttft_ms_q25": 415.621217,
    "ttft_ms_q75": 442.29111650000004,
    "prefill_tokens_per_s": 4601.665926712596,
    "prefill_tokens_per_s_q25": 4594.271032528484,
    "prefill_tokens_per_s_q75": 4908.212707749677,
    "handoff_plus_first_decode_ms": 37.542289,
    "handoff_plus_first_decode_ms_q25": 33.3053835,
    "handoff_plus_first_decode_ms_q75": 42.152677,
    "steady_tpot_ms": 16.547641000000002,
    "steady_tpot_ms_q25": 16.492974500000003,
    "steady_tpot_ms_q75": 16.61902925,
    "steady_tpot_p95_ms": 18.34303435,
    "steady_tpot_p95_ms_q25": 18.236382024999997,
    "steady_tpot_p95_ms_q75": 22.249032649999997,
    "decode_tokens_per_s": 54.51941527970728,
    "decode_tokens_per_s_q25": 52.15687718422036,
    "decode_tokens_per_s_q75": 55.63636080975168,
    "decode_ms": 275.13134400000007,
    "decode_ms_q25": 269.71656299999995,
    "decode_ms_q75": 288.185203,
    "end_to_end_ms": 705.8809909999999,
    "end_to_end_ms_q25": 698.3916389999999,
    "end_to_end_ms_q75": 712.0076795,
    "output_tokens_per_s_e2e": 22.666710400195495,
    "output_tokens_per_s_e2e_q25": 22.473331723265623,
    "output_tokens_per_s_e2e_q75": 22.91241659637302,
    "all_tokens_per_s_e2e": 2901.3389312250233,
    "all_tokens_per_s_e2e_q25": 2876.586460578,
    "all_tokens_per_s_e2e_q75": 2932.7893243357466,
    "energy_net_psys_j": 21.176776688412616,
    "energy_net_psys_j_q25": 20.70068914507565,
    "energy_net_psys_j_q75": 21.378772005881416,
    "prefill_energy_net_psys_j": 11.988368053034002,
    "prefill_energy_net_psys_j_q25": 11.115381583578568,
    "prefill_energy_net_psys_j_q75": 12.034233480063682,
    "decode_energy_net_psys_j": 9.592399270316214,
    "decode_energy_net_psys_j_q25": 9.344538525817736,
    "decode_energy_net_psys_j_q75": 9.787302878965884,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 2048,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 441.579209,
    "ttft_ms_q25": 415.621217,
    "ttft_ms_q75": 442.29111650000004,
    "prefill_tokens_per_s": 4601.665926712596,
    "prefill_tokens_per_s_q25": 4594.271032528484,
    "prefill_tokens_per_s_q75": 4908.212707749677,
    "handoff_plus_first_decode_ms": 31.766454,
    "handoff_plus_first_decode_ms_q25": 31.277967500000003,
    "handoff_plus_first_decode_ms_q75": 67.5615095,
    "steady_tpot_ms": 16.9854185,
    "steady_tpot_ms_q25": 16.967267,
    "steady_tpot_ms_q75": 17.151094750000002,
    "steady_tpot_p95_ms": 17.955736849999997,
    "steady_tpot_p95_ms_q25": 17.881051425,
    "steady_tpot_p95_ms_q75": 19.4568709,
    "decode_tokens_per_s": 54.76064976081041,
    "decode_tokens_per_s_q25": 48.776739515156834,
    "decode_tokens_per_s_q75": 54.95241436007043,
    "decode_ms": 273.919321,
    "decode_ms_q25": 272.966763,
    "decode_ms_q75": 312.222669,
    "end_to_end_ms": 715.4985300000001,
    "end_to_end_ms_q25": 715.2578795000001,
    "end_to_end_ms_q75": 727.843886,
    "output_tokens_per_s_e2e": 22.362030569091452,
    "output_tokens_per_s_e2e_q25": 21.98906212724761,
    "output_tokens_per_s_e2e_q75": 22.369556868625512,
    "all_tokens_per_s_e2e": 2862.339912843706,
    "all_tokens_per_s_e2e_q25": 2814.599952287694,
    "all_tokens_per_s_e2e_q75": 2863.3032791840656,
    "energy_net_psys_j": 20.17528765300994,
    "energy_net_psys_j_q25": 20.033770512601087,
    "energy_net_psys_j_q75": 20.21443138737155,
    "prefill_energy_net_psys_j": 11.988368053034002,
    "prefill_energy_net_psys_j_q25": 11.115381583578568,
    "prefill_energy_net_psys_j_q75": 12.034233480063682,
    "decode_energy_net_psys_j": 8.173476214639802,
    "decode_energy_net_psys_j_q25": 8.038680766899018,
    "decode_energy_net_psys_j_q75": 9.053184376763301,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9865582686801604,
    "energy_reduction_vs_gpu": 0.047291854191892524
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 2048,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 378.987701,
    "ttft_ms_q25": 376.7324235,
    "ttft_ms_q75": 381.44781850000004,
    "prefill_tokens_per_s": 5361.651564518712,
    "prefill_tokens_per_s_q25": 5327.293606253492,
    "prefill_tokens_per_s_q75": 5393.941951970818,
    "handoff_plus_first_decode_ms": 31.766454,
    "handoff_plus_first_decode_ms_q25": 31.277967500000003,
    "handoff_plus_first_decode_ms_q75": 67.5615095,
    "steady_tpot_ms": 16.9854185,
    "steady_tpot_ms_q25": 16.967267,
    "steady_tpot_ms_q75": 17.151094750000002,
    "steady_tpot_p95_ms": 17.955736849999997,
    "steady_tpot_p95_ms_q25": 17.881051425,
    "steady_tpot_p95_ms_q75": 19.4568709,
    "decode_tokens_per_s": 54.76064976081041,
    "decode_tokens_per_s_q25": 48.776739515156834,
    "decode_tokens_per_s_q75": 54.95241436007043,
    "decode_ms": 273.919321,
    "decode_ms_q25": 272.966763,
    "decode_ms_q75": 312.222669,
    "end_to_end_ms": 657.827257,
    "end_to_end_ms_q25": 652.159304,
    "end_to_end_ms_q75": 693.6704875,
    "output_tokens_per_s_e2e": 24.322494742719364,
    "output_tokens_per_s_e2e_q25": 23.127456508498895,
    "output_tokens_per_s_e2e_q75": 24.535736198821198,
    "all_tokens_per_s_e2e": 3113.2793270680786,
    "all_tokens_per_s_e2e_q25": 2960.3144330878586,
    "all_tokens_per_s_e2e_q75": 3140.5742334491133,
    "energy_net_psys_j": 14.641903385774786,
    "energy_net_psys_j_q25": 14.556627562261449,
    "energy_net_psys_j_q75": 15.729274250951343,
    "prefill_energy_net_psys_j": 6.567466419589879,
    "prefill_energy_net_psys_j_q25": 6.517946795362431,
    "prefill_energy_net_psys_j_q75": 6.725609498415488,
    "decode_energy_net_psys_j": 8.173476214639802,
    "decode_energy_net_psys_j_q25": 8.038680766899018,
    "decode_energy_net_psys_j_q75": 9.053184376763301,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0730491682864394,
    "energy_reduction_vs_gpu": 0.30858677875247864
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 2048,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 378.987701,
    "ttft_ms_q25": 376.7324235,
    "ttft_ms_q75": 381.44781850000004,
    "prefill_tokens_per_s": 5361.651564518712,
    "prefill_tokens_per_s_q25": 5327.293606253492,
    "prefill_tokens_per_s_q75": 5393.941951970818,
    "handoff_plus_first_decode_ms": 37.542289,
    "handoff_plus_first_decode_ms_q25": 33.3053835,
    "handoff_plus_first_decode_ms_q75": 42.152677,
    "steady_tpot_ms": 16.547641000000002,
    "steady_tpot_ms_q25": 16.492974500000003,
    "steady_tpot_ms_q75": 16.61902925,
    "steady_tpot_p95_ms": 18.34303435,
    "steady_tpot_p95_ms_q25": 18.236382024999997,
    "steady_tpot_p95_ms_q75": 22.249032649999997,
    "decode_tokens_per_s": 54.51941527970728,
    "decode_tokens_per_s_q25": 52.15687718422036,
    "decode_tokens_per_s_q75": 55.63636080975168,
    "decode_ms": 275.13134400000007,
    "decode_ms_q25": 269.71656299999995,
    "decode_ms_q75": 288.185203,
    "end_to_end_ms": 649.6084900000001,
    "end_to_end_ms_q25": 648.909104,
    "end_to_end_ms_q75": 664.9176265000001,
    "output_tokens_per_s_e2e": 24.63021996525938,
    "output_tokens_per_s_e2e_q25": 24.07589393755326,
    "output_tokens_per_s_e2e_q75": 24.65679474583159,
    "all_tokens_per_s_e2e": 3152.6681555532004,
    "all_tokens_per_s_e2e_q25": 3081.7144240068174,
    "all_tokens_per_s_e2e_q75": 3156.0697274664435,
    "energy_net_psys_j": 16.15986568990609,
    "energy_net_psys_j_q25": 15.862485321180166,
    "energy_net_psys_j_q75": 16.51291237738137,
    "prefill_energy_net_psys_j": 6.567466419589879,
    "prefill_energy_net_psys_j_q25": 6.517946795362431,
    "prefill_energy_net_psys_j_q75": 6.725609498415488,
    "decode_energy_net_psys_j": 9.592399270316214,
    "decode_energy_net_psys_j_q25": 9.344538525817736,
    "decode_energy_net_psys_j_q75": 9.787302878965884,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0866252548515796,
    "energy_reduction_vs_gpu": 0.23690626162439765
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 4096,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1102.167363,
    "ttft_ms_q25": 1054.916456,
    "ttft_ms_q75": 1115.386621,
    "prefill_tokens_per_s": 3701.797147118028,
    "prefill_tokens_per_s_q25": 3658.438333858278,
    "prefill_tokens_per_s_q75": 3875.379813415315,
    "handoff_plus_first_decode_ms": 46.685089999999995,
    "handoff_plus_first_decode_ms_q25": 43.7248115,
    "handoff_plus_first_decode_ms_q75": 48.964158,
    "steady_tpot_ms": 19.858069999999998,
    "steady_tpot_ms_q25": 19.68306975,
    "steady_tpot_ms_q75": 19.95097275,
    "steady_tpot_p95_ms": 23.73208415,
    "steady_tpot_p95_ms_q25": 23.58919285,
    "steady_tpot_p95_ms_q75": 23.794901949999996,
    "decode_tokens_per_s": 44.477808863252505,
    "decode_tokens_per_s_q25": 44.17141662359542,
    "decode_tokens_per_s_q75": 45.31323026243381,
    "decode_ms": 337.24682899999993,
    "decode_ms_q25": 331.14170550000006,
    "decode_ms_q75": 339.6024599999999,
    "end_to_end_ms": 1439.414192,
    "end_to_end_ms_q25": 1394.518916,
    "end_to_end_ms_q75": 1446.5283265,
    "output_tokens_per_s_e2e": 11.115633074152711,
    "output_tokens_per_s_e2e_q25": 11.061233102451435,
    "output_tokens_per_s_e2e_q75": 11.485394984057322,
    "all_tokens_per_s_e2e": 2845.602066983094,
    "all_tokens_per_s_e2e_q25": 2831.6756742275675,
    "all_tokens_per_s_e2e_q75": 2940.2611159186745,
    "energy_net_psys_j": 40.289082094538706,
    "energy_net_psys_j_q25": 39.071873587210305,
    "energy_net_psys_j_q75": 40.4494325020034,
    "prefill_energy_net_psys_j": 29.509258382133034,
    "prefill_energy_net_psys_j_q25": 28.141964137578384,
    "prefill_energy_net_psys_j_q75": 29.925564630126253,
    "decode_energy_net_psys_j": 10.77982371240567,
    "decode_energy_net_psys_j_q25": 10.52386787187714,
    "decode_energy_net_psys_j_q75": 10.929909449631921,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 4096,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1102.167363,
    "ttft_ms_q25": 1054.916456,
    "ttft_ms_q75": 1115.386621,
    "prefill_tokens_per_s": 3701.797147118028,
    "prefill_tokens_per_s_q25": 3658.438333858278,
    "prefill_tokens_per_s_q75": 3875.379813415315,
    "handoff_plus_first_decode_ms": 54.623934,
    "handoff_plus_first_decode_ms_q25": 53.079732500000006,
    "handoff_plus_first_decode_ms_q75": 88.20110700000001,
    "steady_tpot_ms": 19.554413999999998,
    "steady_tpot_ms_q25": 19.4815235,
    "steady_tpot_ms_q75": 19.59206975,
    "steady_tpot_p95_ms": 20.2453457,
    "steady_tpot_p95_ms_q25": 20.175806625,
    "steady_tpot_p95_ms_q75": 23.897542674999997,
    "decode_tokens_per_s": 45.60808645967518,
    "decode_tokens_per_s_q25": 40.88247535057374,
    "decode_tokens_per_s_q75": 45.700504780446224,
    "decode_ms": 328.889045,
    "decode_ms_q25": 328.22528800000003,
    "decode_ms_q75": 371.874015,
    "end_to_end_ms": 1431.056408,
    "end_to_end_ms_q25": 1426.7904709999998,
    "end_to_end_ms_q75": 1443.611909,
    "output_tokens_per_s_e2e": 11.180551591506518,
    "output_tokens_per_s_e2e_q25": 11.084149607968298,
    "output_tokens_per_s_e2e_q75": 11.214080383154492,
    "all_tokens_per_s_e2e": 2862.2212074256686,
    "all_tokens_per_s_e2e_q25": 2837.5422996398843,
    "all_tokens_per_s_e2e_q75": 2870.80457808755,
    "energy_net_psys_j": 38.96763849047304,
    "energy_net_psys_j_q25": 38.37706180776674,
    "energy_net_psys_j_q75": 39.5647979757121,
    "prefill_energy_net_psys_j": 29.509258382133034,
    "prefill_energy_net_psys_j_q25": 28.141964137578384,
    "prefill_energy_net_psys_j_q75": 29.925564630126253,
    "decode_energy_net_psys_j": 9.820086582831687,
    "decode_energy_net_psys_j_q25": 9.639233345585847,
    "decode_energy_net_psys_j_q75": 10.415950907434198,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0058402897001668,
    "energy_reduction_vs_gpu": 0.03279904965233216
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 4096,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 889.630587,
    "ttft_ms_q25": 889.4896515,
    "ttft_ms_q75": 893.698973,
    "prefill_tokens_per_s": 4586.173249459104,
    "prefill_tokens_per_s_q25": 4565.3902227043145,
    "prefill_tokens_per_s_q75": 4586.900022442942,
    "handoff_plus_first_decode_ms": 54.623934,
    "handoff_plus_first_decode_ms_q25": 53.079732500000006,
    "handoff_plus_first_decode_ms_q75": 88.20110700000001,
    "steady_tpot_ms": 19.554413999999998,
    "steady_tpot_ms_q25": 19.4815235,
    "steady_tpot_ms_q75": 19.59206975,
    "steady_tpot_p95_ms": 20.2453457,
    "steady_tpot_p95_ms_q25": 20.175806625,
    "steady_tpot_p95_ms_q75": 23.897542674999997,
    "decode_tokens_per_s": 45.60808645967518,
    "decode_tokens_per_s_q25": 40.88247535057374,
    "decode_tokens_per_s_q75": 45.700504780446224,
    "decode_ms": 328.889045,
    "decode_ms_q25": 328.22528800000003,
    "decode_ms_q75": 371.874015,
    "end_to_end_ms": 1226.656404,
    "end_to_end_ms_q25": 1221.7833255,
    "end_to_end_ms_q75": 1265.572988,
    "output_tokens_per_s_e2e": 13.043587387491435,
    "output_tokens_per_s_e2e_q25": 12.654460578721014,
    "output_tokens_per_s_e2e_q75": 13.095820019616568,
    "all_tokens_per_s_e2e": 3339.1583711978074,
    "all_tokens_per_s_e2e_q25": 3239.5419081525797,
    "all_tokens_per_s_e2e_q75": 3352.5299250218413,
    "energy_net_psys_j": 27.786917847685658,
    "energy_net_psys_j_q25": 27.612264668247487,
    "energy_net_psys_j_q75": 28.929106769659477,
    "prefill_energy_net_psys_j": 17.979231380469308,
    "prefill_energy_net_psys_j_q25": 17.97303132266164,
    "prefill_energy_net_psys_j_q75": 18.51935592003295,
    "decode_energy_net_psys_j": 9.820086582831687,
    "decode_energy_net_psys_j_q25": 9.639233345585847,
    "decode_energy_net_psys_j_q75": 10.415950907434198,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.1734453000092109,
    "energy_reduction_vs_gpu": 0.31031146893633876
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 4096,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 889.630587,
    "ttft_ms_q25": 889.4896515,
    "ttft_ms_q75": 893.698973,
    "prefill_tokens_per_s": 4586.173249459104,
    "prefill_tokens_per_s_q25": 4565.3902227043145,
    "prefill_tokens_per_s_q75": 4586.900022442942,
    "handoff_plus_first_decode_ms": 46.685089999999995,
    "handoff_plus_first_decode_ms_q25": 43.7248115,
    "handoff_plus_first_decode_ms_q75": 48.964158,
    "steady_tpot_ms": 19.858069999999998,
    "steady_tpot_ms_q25": 19.68306975,
    "steady_tpot_ms_q75": 19.95097275,
    "steady_tpot_p95_ms": 23.73208415,
    "steady_tpot_p95_ms_q25": 23.58919285,
    "steady_tpot_p95_ms_q75": 23.794901949999996,
    "decode_tokens_per_s": 44.477808863252505,
    "decode_tokens_per_s_q25": 44.17141662359542,
    "decode_tokens_per_s_q75": 45.31323026243381,
    "decode_ms": 337.24682899999993,
    "decode_ms_q25": 331.14170550000006,
    "decode_ms_q75": 339.6024599999999,
    "end_to_end_ms": 1231.5886779999998,
    "end_to_end_ms_q25": 1222.986988,
    "end_to_end_ms_q75": 1233.301433,
    "output_tokens_per_s_e2e": 12.991350347571158,
    "output_tokens_per_s_e2e_q25": 12.973333550289995,
    "output_tokens_per_s_e2e_q75": 13.083370210258263,
    "all_tokens_per_s_e2e": 3325.7856889782165,
    "all_tokens_per_s_e2e_q25": 3321.1733888742388,
    "all_tokens_per_s_e2e_q75": 3349.3427738261153,
    "energy_net_psys_j": 28.75905509287498,
    "energy_net_psys_j_q25": 28.496899194538784,
    "energy_net_psys_j_q75": 29.44926536966487,
    "prefill_energy_net_psys_j": 17.979231380469308,
    "prefill_energy_net_psys_j_q25": 17.97303132266164,
    "prefill_energy_net_psys_j_q75": 18.51935592003295,
    "decode_energy_net_psys_j": 10.77982371240567,
    "decode_energy_net_psys_j_q25": 10.52386787187714,
    "decode_energy_net_psys_j_q75": 10.929909449631921,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.1687458789711285,
    "energy_reduction_vs_gpu": 0.2861824197088534
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 8192,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2856.822478,
    "ttft_ms_q25": 2751.7923769999998,
    "ttft_ms_q75": 2864.265589,
    "prefill_tokens_per_s": 2861.920914919376,
    "prefill_tokens_per_s_q25": 2854.5031726382854,
    "prefill_tokens_per_s_q75": 2975.4890451113583,
    "handoff_plus_first_decode_ms": 54.975293,
    "handoff_plus_first_decode_ms_q25": 54.7470285,
    "handoff_plus_first_decode_ms_q75": 62.376990000000006,
    "steady_tpot_ms": 22.787651,
    "steady_tpot_ms_q25": 22.78760075,
    "steady_tpot_ms_q75": 22.823230000000002,
    "steady_tpot_p95_ms": 26.0898805,
    "steady_tpot_p95_ms_q25": 25.599848625,
    "steady_tpot_p95_ms_q75": 26.176459174999998,
    "decode_tokens_per_s": 39.09940484893042,
    "decode_tokens_per_s_q25": 38.331621450663874,
    "decode_tokens_per_s_q75": 39.24716908652915,
    "decode_ms": 383.63755299999957,
    "decode_ms_q25": 382.1985884999999,
    "decode_ms_q75": 391.47888549999993,
    "end_to_end_ms": 3237.5821020000003,
    "end_to_end_ms_q25": 3141.8322980000003,
    "end_to_end_ms_q75": 3246.4641775,
    "output_tokens_per_s_e2e": 4.941959615515566,
    "output_tokens_per_s_e2e_q25": 4.928475686307573,
    "output_tokens_per_s_e2e_q75": 5.0973039522147845,
    "all_tokens_per_s_e2e": 2530.28332314397,
    "all_tokens_per_s_e2e_q25": 2523.3795513894775,
    "all_tokens_per_s_e2e_q75": 2609.8196235339697,
    "energy_net_psys_j": 89.21420497592791,
    "energy_net_psys_j_q25": 87.10804362204142,
    "energy_net_psys_j_q75": 90.54855250851622,
    "prefill_energy_net_psys_j": 78.6067166446301,
    "prefill_energy_net_psys_j_q25": 76.10534310695107,
    "prefill_energy_net_psys_j_q75": 79.3695969585709,
    "decode_energy_net_psys_j": 11.397912698882905,
    "decode_energy_net_psys_j_q25": 11.002700515090362,
    "decode_energy_net_psys_j_q75": 11.57416773373787,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 8192,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2856.822478,
    "ttft_ms_q25": 2751.7923769999998,
    "ttft_ms_q75": 2864.265589,
    "prefill_tokens_per_s": 2861.920914919376,
    "prefill_tokens_per_s_q25": 2854.5031726382854,
    "prefill_tokens_per_s_q75": 2975.4890451113583,
    "handoff_plus_first_decode_ms": 87.489434,
    "handoff_plus_first_decode_ms_q25": 81.8039895,
    "handoff_plus_first_decode_ms_q75": 116.94190849999998,
    "steady_tpot_ms": 22.611179500000002,
    "steady_tpot_ms_q25": 22.605743000000004,
    "steady_tpot_ms_q75": 22.923969,
    "steady_tpot_p95_ms": 24.377819849999995,
    "steady_tpot_p95_ms_q25": 24.140172874999998,
    "steady_tpot_p95_ms_q75": 27.619299924999993,
    "decode_tokens_per_s": 36.81603411132208,
    "decode_tokens_per_s_q25": 33.677503954482,
    "decode_tokens_per_s_q75": 37.2919916797083,
    "decode_ms": 407.4311740000003,
    "decode_ms_q25": 402.2966630000003,
    "decode_ms_q75": 449.3034069999999,
    "end_to_end_ms": 3264.2536520000003,
    "end_to_end_ms_q25": 3201.095784,
    "end_to_end_ms_q75": 3266.5622520000006,
    "output_tokens_per_s_e2e": 4.901579872690604,
    "output_tokens_per_s_e2e_q25": 4.898118191943207,
    "output_tokens_per_s_e2e_q75": 5.000234894835042,
    "all_tokens_per_s_e2e": 2509.6088948175893,
    "all_tokens_per_s_e2e_q25": 2507.836514274922,
    "all_tokens_per_s_e2e_q75": 2560.1202661555417,
    "energy_net_psys_j": 88.81601267346514,
    "energy_net_psys_j_q25": 87.01605052325792,
    "energy_net_psys_j_q75": 89.57725471644027,
    "prefill_energy_net_psys_j": 78.6067166446301,
    "prefill_energy_net_psys_j_q25": 76.10534310695107,
    "prefill_energy_net_psys_j_q75": 79.3695969585709,
    "decode_energy_net_psys_j": 10.209296028835045,
    "decode_energy_net_psys_j_q25": 10.207657757869368,
    "decode_energy_net_psys_j_q75": 10.910707416306865,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9918292042091587,
    "energy_reduction_vs_gpu": 0.0044633284864245
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 8192,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2978.993062,
    "ttft_ms_q25": 2976.052053,
    "ttft_ms_q75": 2985.345813,
    "prefill_tokens_per_s": 2744.551541355688,
    "prefill_tokens_per_s_q25": 2738.723597070557,
    "prefill_tokens_per_s_q75": 2747.2664587148565,
    "handoff_plus_first_decode_ms": 87.489434,
    "handoff_plus_first_decode_ms_q25": 81.8039895,
    "handoff_plus_first_decode_ms_q75": 116.94190849999998,
    "steady_tpot_ms": 22.611179500000002,
    "steady_tpot_ms_q25": 22.605743000000004,
    "steady_tpot_ms_q75": 22.923969,
    "steady_tpot_p95_ms": 24.377819849999995,
    "steady_tpot_p95_ms_q25": 24.140172874999998,
    "steady_tpot_p95_ms_q75": 27.619299924999993,
    "decode_tokens_per_s": 36.81603411132208,
    "decode_tokens_per_s_q25": 33.677503954482,
    "decode_tokens_per_s_q75": 37.2919916797083,
    "decode_ms": 407.4311740000003,
    "decode_ms_q25": 402.2966630000003,
    "decode_ms_q75": 449.3034069999999,
    "end_to_end_ms": 3388.860716,
    "end_to_end_ms_q25": 3384.701467,
    "end_to_end_ms_q75": 3429.514709,
    "output_tokens_per_s_e2e": 4.7213507254690015,
    "output_tokens_per_s_e2e_q25": 4.666038786532679,
    "output_tokens_per_s_e2e_q75": 4.727159637772772,
    "all_tokens_per_s_e2e": 2417.3315714401288,
    "all_tokens_per_s_e2e_q25": 2389.011858704732,
    "all_tokens_per_s_e2e_q75": 2420.305734539659,
    "energy_net_psys_j": 64.2541216746021,
    "energy_net_psys_j_q25": 63.08175203277463,
    "energy_net_psys_j_q75": 66.60504801560094,
    "prefill_energy_net_psys_j": 54.04482564576706,
    "prefill_energy_net_psys_j_q25": 52.87409427490526,
    "prefill_energy_net_psys_j_q75": 55.69434059929409,
    "decode_energy_net_psys_j": 10.209296028835045,
    "decode_energy_net_psys_j_q25": 10.207657757869368,
    "decode_energy_net_psys_j_q75": 10.910707416306865,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.9553600378776972,
    "energy_reduction_vs_gpu": 0.27977700757475366
  },
  {
    "precision": "FP16",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 8192,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2978.993062,
    "ttft_ms_q25": 2976.052053,
    "ttft_ms_q75": 2985.345813,
    "prefill_tokens_per_s": 2744.551541355688,
    "prefill_tokens_per_s_q25": 2738.723597070557,
    "prefill_tokens_per_s_q75": 2747.2664587148565,
    "handoff_plus_first_decode_ms": 54.975293,
    "handoff_plus_first_decode_ms_q25": 54.7470285,
    "handoff_plus_first_decode_ms_q75": 62.376990000000006,
    "steady_tpot_ms": 22.787651,
    "steady_tpot_ms_q25": 22.78760075,
    "steady_tpot_ms_q75": 22.823230000000002,
    "steady_tpot_p95_ms": 26.0898805,
    "steady_tpot_p95_ms_q25": 25.599848625,
    "steady_tpot_p95_ms_q75": 26.176459174999998,
    "decode_tokens_per_s": 39.09940484893042,
    "decode_tokens_per_s_q25": 38.331621450663874,
    "decode_tokens_per_s_q75": 39.24716908652915,
    "decode_ms": 383.63755299999957,
    "decode_ms_q25": 382.1985884999999,
    "decode_ms_q75": 391.47888549999993,
    "end_to_end_ms": 3375.3361169999994,
    "end_to_end_ms_q25": 3364.6033924999997,
    "end_to_end_ms_q75": 3376.8246984999996,
    "output_tokens_per_s_e2e": 4.740268656331865,
    "output_tokens_per_s_e2e_q25": 4.738179958321345,
    "output_tokens_per_s_e2e_q75": 4.755437994264247,
    "all_tokens_per_s_e2e": 2427.0175520419148,
    "all_tokens_per_s_e2e_q25": 2425.9481386605285,
    "all_tokens_per_s_e2e_q75": 2434.7842530632943,
    "energy_net_psys_j": 64.65231397706488,
    "energy_net_psys_j_q25": 64.05304982485059,
    "energy_net_psys_j_q75": 66.69704111438445,
    "prefill_energy_net_psys_j": 54.04482564576706,
    "prefill_energy_net_psys_j_q25": 52.87409427490526,
    "prefill_energy_net_psys_j_q75": 55.69434059929409,
    "decode_energy_net_psys_j": 11.397912698882905,
    "decode_energy_net_psys_j_q25": 11.002700515090362,
    "decode_energy_net_psys_j_q75": 11.57416773373787,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9591880600257272,
    "energy_reduction_vs_gpu": 0.27531367908832916
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 1024,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 360.493108,
    "ttft_ms_q25": 347.1690065,
    "ttft_ms_q75": 369.713306,
    "prefill_tokens_per_s": 2796.169961729199,
    "prefill_tokens_per_s_q25": 2728.133642911648,
    "prefill_tokens_per_s_q75": 2907.768047616809,
    "handoff_plus_first_decode_ms": 55.536018999999996,
    "handoff_plus_first_decode_ms_q25": 54.2231775,
    "handoff_plus_first_decode_ms_q75": 57.620129999999996,
    "steady_tpot_ms": 43.521603,
    "steady_tpot_ms_q25": 43.449562,
    "steady_tpot_ms_q75": 43.5519585,
    "steady_tpot_p95_ms": 45.06358405,
    "steady_tpot_p95_ms_q25": 44.741105399999995,
    "steady_tpot_p95_ms_q75": 45.35408905,
    "decode_tokens_per_s": 22.477790538086463,
    "decode_tokens_per_s_q25": 22.3496970718926,
    "decode_tokens_per_s_q75": 22.51589054630978,
    "decode_ms": 667.3253749999998,
    "decode_ms_q25": 666.1980755,
    "decode_ms_q75": 671.1720829999999,
    "end_to_end_ms": 1027.8184829999998,
    "end_to_end_ms_q25": 1018.3410895,
    "end_to_end_ms_q75": 1035.9113815,
    "output_tokens_per_s_e2e": 15.566951037209556,
    "output_tokens_per_s_e2e_q25": 15.446279353087142,
    "output_tokens_per_s_e2e_q75": 15.713188949387204,
    "all_tokens_per_s_e2e": 996.2848663814116,
    "all_tokens_per_s_e2e_q25": 988.5618785975771,
    "all_tokens_per_s_e2e_q75": 1005.6440927607811,
    "energy_net_psys_j": 34.01359510119007,
    "energy_net_psys_j_q25": 33.44256860175318,
    "energy_net_psys_j_q75": 34.20331383758705,
    "prefill_energy_net_psys_j": 11.490362435506256,
    "prefill_energy_net_psys_j_q25": 11.080950246895839,
    "prefill_energy_net_psys_j_q75": 11.726210388944478,
    "decode_energy_net_psys_j": 22.43097423160133,
    "decode_energy_net_psys_j_q25": 21.906076949205676,
    "decode_energy_net_psys_j_q75": 22.88651563725299,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 1024,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 360.493108,
    "ttft_ms_q25": 347.1690065,
    "ttft_ms_q75": 369.713306,
    "prefill_tokens_per_s": 2796.169961729199,
    "prefill_tokens_per_s_q25": 2728.133642911648,
    "prefill_tokens_per_s_q75": 2907.768047616809,
    "handoff_plus_first_decode_ms": 52.85397,
    "handoff_plus_first_decode_ms_q25": 52.370791499999996,
    "handoff_plus_first_decode_ms_q75": 89.395224,
    "steady_tpot_ms": 47.742450999999996,
    "steady_tpot_ms_q25": 47.442558749999996,
    "steady_tpot_ms_q75": 48.24036325,
    "steady_tpot_p95_ms": 49.104597749999996,
    "steady_tpot_p95_ms_q25": 48.51265959999999,
    "steady_tpot_p95_ms_q75": 49.31260585,
    "decode_tokens_per_s": 20.557439560377347,
    "decode_tokens_per_s_q25": 19.68496873840741,
    "decode_tokens_per_s_q75": 20.773227416650087,
    "decode_ms": 729.6628529999998,
    "decode_ms_q25": 722.1611964999998,
    "decode_ms_q75": 763.5025669999999,
    "end_to_end_ms": 1093.5930439999997,
    "end_to_end_ms_q25": 1091.8745024999998,
    "end_to_end_ms_q75": 1112.390115,
    "output_tokens_per_s_e2e": 14.630670968313149,
    "output_tokens_per_s_e2e_q25": 14.387551381764876,
    "output_tokens_per_s_e2e_q75": 14.653735021651741,
    "all_tokens_per_s_e2e": 936.3629419720415,
    "all_tokens_per_s_e2e_q25": 920.8032884329521,
    "all_tokens_per_s_e2e_q75": 937.8390413857114,
    "energy_net_psys_j": 27.41282792159977,
    "energy_net_psys_j_q25": 26.87561944603535,
    "energy_net_psys_j_q75": 27.501389661654663,
    "prefill_energy_net_psys_j": 11.490362435506256,
    "prefill_energy_net_psys_j_q25": 11.080950246895839,
    "prefill_energy_net_psys_j_q75": 11.726210388944478,
    "decode_energy_net_psys_j": 15.627893059326851,
    "decode_energy_net_psys_j_q25": 15.237970797145763,
    "decode_energy_net_psys_j_q75": 16.1845914613206,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9398546274952349,
    "energy_reduction_vs_gpu": 0.19406261407984338
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 1024,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 470.919173,
    "ttft_ms_q25": 470.917144,
    "ttft_ms_q75": 473.960953,
    "prefill_tokens_per_s": 2140.494712879316,
    "prefill_tokens_per_s_q25": 2126.845076184798,
    "prefill_tokens_per_s_q75": 2140.5039354836326,
    "handoff_plus_first_decode_ms": 52.85397,
    "handoff_plus_first_decode_ms_q25": 52.370791499999996,
    "handoff_plus_first_decode_ms_q75": 89.395224,
    "steady_tpot_ms": 47.742450999999996,
    "steady_tpot_ms_q25": 47.442558749999996,
    "steady_tpot_ms_q75": 48.24036325,
    "steady_tpot_p95_ms": 49.104597749999996,
    "steady_tpot_p95_ms_q25": 48.51265959999999,
    "steady_tpot_p95_ms_q75": 49.31260585,
    "decode_tokens_per_s": 20.557439560377347,
    "decode_tokens_per_s_q25": 19.68496873840741,
    "decode_tokens_per_s_q75": 20.773227416650087,
    "decode_ms": 729.6628529999998,
    "decode_ms_q25": 722.1611964999998,
    "decode_ms_q75": 763.5025669999999,
    "end_to_end_ms": 1206.6655859999998,
    "end_to_end_ms_q25": 1196.1221495,
    "end_to_end_ms_q75": 1237.461491,
    "output_tokens_per_s_e2e": 13.259680383393317,
    "output_tokens_per_s_e2e_q25": 12.937707999313213,
    "output_tokens_per_s_e2e_q75": 13.377599671754059,
    "all_tokens_per_s_e2e": 848.6195445371723,
    "all_tokens_per_s_e2e_q25": 828.0133119560456,
    "all_tokens_per_s_e2e_q75": 856.1663789922598,
    "energy_net_psys_j": 24.786369496670208,
    "energy_net_psys_j_q25": 23.823591376537124,
    "energy_net_psys_j_q75": 25.291509651087466,
    "prefill_energy_net_psys_j": 9.055359942190375,
    "prefill_energy_net_psys_j_q25": 8.534062331814866,
    "prefill_energy_net_psys_j_q75": 9.106918189766866,
    "decode_energy_net_psys_j": 15.627893059326851,
    "decode_energy_net_psys_j_q25": 15.237970797145763,
    "decode_energy_net_psys_j_q75": 16.1845914613206,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.851784036045261,
    "energy_reduction_vs_gpu": 0.2712805152489458
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 1024,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 470.919173,
    "ttft_ms_q25": 470.917144,
    "ttft_ms_q75": 473.960953,
    "prefill_tokens_per_s": 2140.494712879316,
    "prefill_tokens_per_s_q25": 2126.845076184798,
    "prefill_tokens_per_s_q75": 2140.5039354836326,
    "handoff_plus_first_decode_ms": 55.536018999999996,
    "handoff_plus_first_decode_ms_q25": 54.2231775,
    "handoff_plus_first_decode_ms_q75": 57.620129999999996,
    "steady_tpot_ms": 43.521603,
    "steady_tpot_ms_q25": 43.449562,
    "steady_tpot_ms_q75": 43.5519585,
    "steady_tpot_p95_ms": 45.06358405,
    "steady_tpot_p95_ms_q25": 44.741105399999995,
    "steady_tpot_p95_ms_q75": 45.35408905,
    "decode_tokens_per_s": 22.477790538086463,
    "decode_tokens_per_s_q25": 22.3496970718926,
    "decode_tokens_per_s_q75": 22.51589054630978,
    "decode_ms": 667.3253749999998,
    "decode_ms_q25": 666.1980755,
    "decode_ms_q75": 671.1720829999999,
    "end_to_end_ms": 1144.3281079999997,
    "end_to_end_ms_q25": 1140.1590285,
    "end_to_end_ms_q75": 1145.131007,
    "output_tokens_per_s_e2e": 13.982003839758873,
    "output_tokens_per_s_e2e_q25": 13.972207343781083,
    "output_tokens_per_s_e2e_q75": 14.033317749383311,
    "all_tokens_per_s_e2e": 894.8482457445679,
    "all_tokens_per_s_e2e_q25": 894.2212700019893,
    "all_tokens_per_s_e2e_q75": 898.1323359605319,
    "energy_net_psys_j": 31.589450668944686,
    "energy_net_psys_j_q25": 30.491697528597037,
    "energy_net_psys_j_q75": 31.993433827019853,
    "prefill_energy_net_psys_j": 9.055359942190375,
    "prefill_energy_net_psys_j_q25": 8.534062331814866,
    "prefill_energy_net_psys_j_q75": 9.106918189766866,
    "decode_energy_net_psys_j": 22.43097423160133,
    "decode_energy_net_psys_j_q25": 21.906076949205676,
    "decode_energy_net_psys_j_q75": 22.88651563725299,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.898185123492571,
    "energy_reduction_vs_gpu": 0.07126986797583679
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 2048,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 770.339746,
    "ttft_ms_q25": 747.4911930000001,
    "ttft_ms_q75": 774.9201585000001,
    "prefill_tokens_per_s": 2637.797167485111,
    "prefill_tokens_per_s_q25": 2622.297244113774,
    "prefill_tokens_per_s_q75": 2720.9689890943573,
    "handoff_plus_first_decode_ms": 65.452983,
    "handoff_plus_first_decode_ms_q25": 63.2263055,
    "handoff_plus_first_decode_ms_q75": 67.77745300000001,
    "steady_tpot_ms": 46.5253935,
    "steady_tpot_ms_q25": 46.489628499999995,
    "steady_tpot_ms_q75": 46.530795499999996,
    "steady_tpot_p95_ms": 53.8247413,
    "steady_tpot_p95_ms_q25": 51.810425925,
    "steady_tpot_p95_ms_q75": 54.267149575000005,
    "decode_tokens_per_s": 20.416328324438073,
    "decode_tokens_per_s_q25": 20.34733433084017,
    "decode_tokens_per_s_q75": 20.630648167284473,
    "decode_ms": 734.706053,
    "decode_ms_q25": 727.1520915000001,
    "decode_ms_q75": 737.2057794999998,
    "end_to_end_ms": 1499.0987010000001,
    "end_to_end_ms_q25": 1481.7234235,
    "end_to_end_ms_q75": 1502.0722500000002,
    "output_tokens_per_s_e2e": 10.67307975740818,
    "output_tokens_per_s_e2e_q25": 10.651992740879749,
    "output_tokens_per_s_e2e_q75": 10.799721583036305,
    "all_tokens_per_s_e2e": 1366.154208948247,
    "all_tokens_per_s_e2e_q25": 1363.4550708326078,
    "all_tokens_per_s_e2e_q75": 1382.364362628647,
    "energy_net_psys_j": 48.04695553438617,
    "energy_net_psys_j_q25": 47.793525539450656,
    "energy_net_psys_j_q75": 48.689665940820746,
    "prefill_energy_net_psys_j": 25.69547178335022,
    "prefill_energy_net_psys_j_q25": 24.8980657021562,
    "prefill_energy_net_psys_j_q75": 25.883410347253655,
    "decode_energy_net_psys_j": 23.261027436098225,
    "decode_energy_net_psys_j_q25": 22.806255593567087,
    "decode_energy_net_psys_j_q75": 23.35023167982559,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 2048,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 770.339746,
    "ttft_ms_q25": 747.4911930000001,
    "ttft_ms_q75": 774.9201585000001,
    "prefill_tokens_per_s": 2637.797167485111,
    "prefill_tokens_per_s_q25": 2622.297244113774,
    "prefill_tokens_per_s_q75": 2720.9689890943573,
    "handoff_plus_first_decode_ms": 90.44789899999999,
    "handoff_plus_first_decode_ms_q25": 83.576206,
    "handoff_plus_first_decode_ms_q75": 114.51614899999998,
    "steady_tpot_ms": 49.289752500000006,
    "steady_tpot_ms_q25": 49.0961465,
    "steady_tpot_ms_q75": 49.381530500000004,
    "steady_tpot_p95_ms": 50.79916185,
    "steady_tpot_p95_ms_q25": 50.251029375,
    "steady_tpot_p95_ms_q75": 51.060063325,
    "decode_tokens_per_s": 19.106795571571624,
    "decode_tokens_per_s_q25": 18.54293616734449,
    "decode_tokens_per_s_q75": 19.38125119236078,
    "decode_ms": 785.0609979999999,
    "decode_ms_q25": 774.099072,
    "decode_ms_q75": 809.6820634999999,
    "end_to_end_ms": 1555.400744,
    "end_to_end_ms_q25": 1549.0192305,
    "end_to_end_ms_q75": 1557.1732565,
    "output_tokens_per_s_e2e": 10.286738039518387,
    "output_tokens_per_s_e2e_q25": 10.27504207669412,
    "output_tokens_per_s_e2e_q75": 10.32929174927586,
    "all_tokens_per_s_e2e": 1316.7024690583535,
    "all_tokens_per_s_e2e_q25": 1315.2053858168474,
    "all_tokens_per_s_e2e_q75": 1322.1493439073101,
    "energy_net_psys_j": 43.01036802874478,
    "energy_net_psys_j_q25": 42.28801534424129,
    "energy_net_psys_j_q75": 43.67669184408312,
    "prefill_energy_net_psys_j": 25.69547178335022,
    "prefill_energy_net_psys_j_q25": 24.8980657021562,
    "prefill_energy_net_psys_j_q75": 25.883410347253655,
    "decode_energy_net_psys_j": 17.46500303877562,
    "decode_energy_net_psys_j_q25": 17.389949642085092,
    "decode_energy_net_psys_j_q75": 17.86833489352,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9638022270355813,
    "energy_reduction_vs_gpu": 0.10482636099672982
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 2048,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 953.900271,
    "ttft_ms_q25": 950.9068265,
    "ttft_ms_q75": 958.8231645000001,
    "prefill_tokens_per_s": 2130.201722104344,
    "prefill_tokens_per_s_q25": 2119.3204778965087,
    "prefill_tokens_per_s_q75": 2136.9287509131045,
    "handoff_plus_first_decode_ms": 90.44789899999999,
    "handoff_plus_first_decode_ms_q25": 83.576206,
    "handoff_plus_first_decode_ms_q75": 114.51614899999998,
    "steady_tpot_ms": 49.289752500000006,
    "steady_tpot_ms_q25": 49.0961465,
    "steady_tpot_ms_q75": 49.381530500000004,
    "steady_tpot_p95_ms": 50.79916185,
    "steady_tpot_p95_ms_q25": 50.251029375,
    "steady_tpot_p95_ms_q75": 51.060063325,
    "decode_tokens_per_s": 19.106795571571624,
    "decode_tokens_per_s_q25": 18.54293616734449,
    "decode_tokens_per_s_q75": 19.38125119236078,
    "decode_ms": 785.0609979999999,
    "decode_ms_q25": 774.099072,
    "decode_ms_q75": 809.6820634999999,
    "end_to_end_ms": 1732.9743799999999,
    "end_to_end_ms_q25": 1725.0058985,
    "end_to_end_ms_q75": 1765.5117834999999,
    "output_tokens_per_s_e2e": 9.232681212517406,
    "output_tokens_per_s_e2e_q25": 9.06560709898897,
    "output_tokens_per_s_e2e_q75": 9.275528531223998,
    "all_tokens_per_s_e2e": 1181.783195202228,
    "all_tokens_per_s_e2e_q25": 1160.3977086705881,
    "all_tokens_per_s_e2e_q75": 1187.2676519966717,
    "energy_net_psys_j": 39.48190518484798,
    "energy_net_psys_j_q25": 39.055324441821085,
    "energy_net_psys_j_q75": 39.8651912459966,
    "prefill_energy_net_psys_j": 21.2102384365836,
    "prefill_energy_net_psys_j_q25": 21.186989548301078,
    "prefill_energy_net_psys_j_q75": 22.071909749167126,
    "decode_energy_net_psys_j": 17.46500303877562,
    "decode_energy_net_psys_j_q25": 17.389949642085092,
    "decode_energy_net_psys_j_q75": 17.86833489352,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.8650437757769969,
    "energy_reduction_vs_gpu": 0.1782641637597281
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 2048,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 953.900271,
    "ttft_ms_q25": 950.9068265,
    "ttft_ms_q75": 958.8231645000001,
    "prefill_tokens_per_s": 2130.201722104344,
    "prefill_tokens_per_s_q25": 2119.3204778965087,
    "prefill_tokens_per_s_q75": 2136.9287509131045,
    "handoff_plus_first_decode_ms": 65.452983,
    "handoff_plus_first_decode_ms_q25": 63.2263055,
    "handoff_plus_first_decode_ms_q75": 67.77745300000001,
    "steady_tpot_ms": 46.5253935,
    "steady_tpot_ms_q25": 46.489628499999995,
    "steady_tpot_ms_q75": 46.530795499999996,
    "steady_tpot_p95_ms": 53.8247413,
    "steady_tpot_p95_ms_q25": 51.810425925,
    "steady_tpot_p95_ms_q75": 54.267149575000005,
    "decode_tokens_per_s": 20.416328324438073,
    "decode_tokens_per_s_q25": 20.34733433084017,
    "decode_tokens_per_s_q75": 20.630648167284473,
    "decode_ms": 734.706053,
    "decode_ms_q25": 727.1520915000001,
    "decode_ms_q75": 737.2057794999998,
    "end_to_end_ms": 1682.619435,
    "end_to_end_ms_q25": 1678.0589180000002,
    "end_to_end_ms_q75": 1693.0354995,
    "output_tokens_per_s_e2e": 9.508983236010227,
    "output_tokens_per_s_e2e_q25": 9.450838828027694,
    "output_tokens_per_s_e2e_q75": 9.534896544128493,
    "all_tokens_per_s_e2e": 1217.149854209309,
    "all_tokens_per_s_e2e_q25": 1209.7073699875448,
    "all_tokens_per_s_e2e_q75": 1220.4667576484471,
    "energy_net_psys_j": 44.60317658357152,
    "energy_net_psys_j_q25": 44.53722122812667,
    "energy_net_psys_j_q75": 44.94412069817906,
    "prefill_energy_net_psys_j": 21.2102384365836,
    "prefill_energy_net_psys_j_q25": 21.186989548301078,
    "prefill_energy_net_psys_j_q75": 22.071909749167126,
    "decode_energy_net_psys_j": 23.261027436098225,
    "decode_energy_net_psys_j_q25": 22.806255593567087,
    "decode_energy_net_psys_j_q75": 23.35023167982559,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.8909315260583568,
    "energy_reduction_vs_gpu": 0.07167527916206917
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 4096,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1809.143317,
    "ttft_ms_q25": 1753.3364430000001,
    "ttft_ms_q75": 1820.160491,
    "prefill_tokens_per_s": 2255.211050258657,
    "prefill_tokens_per_s_q25": 2241.6427040395465,
    "prefill_tokens_per_s_q75": 2329.3518965941544,
    "handoff_plus_first_decode_ms": 72.596193,
    "handoff_plus_first_decode_ms_q25": 68.940962,
    "handoff_plus_first_decode_ms_q75": 74.4178035,
    "steady_tpot_ms": 44.69485,
    "steady_tpot_ms_q25": 44.68039125,
    "steady_tpot_ms_q75": 44.700368,
    "steady_tpot_p95_ms": 47.106267450000004,
    "steady_tpot_p95_ms_q25": 46.839246125,
    "steady_tpot_p95_ms_q75": 47.156936875,
    "decode_tokens_per_s": 21.282415866569966,
    "decode_tokens_per_s_q25": 21.25051559404212,
    "decode_tokens_per_s_q75": 21.39729160432636,
    "decode_ms": 704.8072030000001,
    "decode_ms_q25": 701.0435074999999,
    "decode_ms_q75": 705.8668170000001,
    "end_to_end_ms": 2506.423129,
    "end_to_end_ms_q25": 2455.4395645,
    "end_to_end_ms_q75": 2521.2039985,
    "output_tokens_per_s_e2e": 6.383598928239862,
    "output_tokens_per_s_e2e_q25": 6.346392419680106,
    "output_tokens_per_s_e2e_q75": 6.51895537797321,
    "all_tokens_per_s_e2e": 1634.2013256294047,
    "all_tokens_per_s_e2e_q25": 1624.676459438107,
    "all_tokens_per_s_e2e_q75": 1668.8525767611418,
    "energy_net_psys_j": 81.4949949933806,
    "energy_net_psys_j_q25": 78.42831116230036,
    "energy_net_psys_j_q75": 82.18184736906498,
    "prefill_energy_net_psys_j": 59.12704217341829,
    "prefill_energy_net_psys_j_q25": 56.45040936864072,
    "prefill_energy_net_psys_j_q75": 60.17508669734304,
    "decode_energy_net_psys_j": 21.645568523481565,
    "decode_energy_net_psys_j_q25": 21.616709645419277,
    "decode_energy_net_psys_j_q75": 22.006760671721942,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 4096,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1809.143317,
    "ttft_ms_q25": 1753.3364430000001,
    "ttft_ms_q75": 1820.160491,
    "prefill_tokens_per_s": 2255.211050258657,
    "prefill_tokens_per_s_q25": 2241.6427040395465,
    "prefill_tokens_per_s_q75": 2329.3518965941544,
    "handoff_plus_first_decode_ms": 128.9693,
    "handoff_plus_first_decode_ms_q25": 125.140123,
    "handoff_plus_first_decode_ms_q75": 151.3195595,
    "steady_tpot_ms": 50.9610385,
    "steady_tpot_ms_q25": 50.9127735,
    "steady_tpot_ms_q75": 51.0554525,
    "steady_tpot_p95_ms": 51.880345350000006,
    "steady_tpot_p95_ms_q25": 51.671559775000006,
    "steady_tpot_p95_ms_q75": 54.092308575000004,
    "decode_tokens_per_s": 17.818719397930845,
    "decode_tokens_per_s_q25": 17.223284756227333,
    "decode_tokens_per_s_q75": 17.871784564224235,
    "decode_ms": 841.8113370000001,
    "decode_ms_q25": 839.3192180000001,
    "decode_ms_q75": 871.956162,
    "end_to_end_ms": 2645.970416,
    "end_to_end_ms_q25": 2622.800486,
    "end_to_end_ms_q75": 2659.479709,
    "output_tokens_per_s_e2e": 6.046930798337391,
    "output_tokens_per_s_e2e_q25": 6.016369595188279,
    "output_tokens_per_s_e2e_q75": 6.100825750060032,
    "all_tokens_per_s_e2e": 1548.0142843743722,
    "all_tokens_per_s_e2e_q25": 1540.1906163681995,
    "all_tokens_per_s_e2e_q75": 1561.8113920153683,
    "energy_net_psys_j": 78.1824820464143,
    "energy_net_psys_j_q25": 75.37543617265496,
    "energy_net_psys_j_q75": 79.17943603316526,
    "prefill_energy_net_psys_j": 59.12704217341829,
    "prefill_energy_net_psys_j_q25": 56.45040936864072,
    "prefill_energy_net_psys_j_q75": 60.17508669734304,
    "decode_energy_net_psys_j": 18.953258798648438,
    "decode_energy_net_psys_j_q25": 18.873936266840467,
    "decode_energy_net_psys_j_q75": 19.004349335822223,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9472604507759544,
    "energy_reduction_vs_gpu": 0.04064682680495124
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 4096,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2212.2164030000004,
    "ttft_ms_q25": 2208.340892,
    "ttft_ms_q75": 2226.989219,
    "prefill_tokens_per_s": 1844.3041984803508,
    "prefill_tokens_per_s_q25": 1832.1505594469477,
    "prefill_tokens_per_s_q75": 1847.5465361782622,
    "handoff_plus_first_decode_ms": 128.9693,
    "handoff_plus_first_decode_ms_q25": 125.140123,
    "handoff_plus_first_decode_ms_q75": 151.3195595,
    "steady_tpot_ms": 50.9610385,
    "steady_tpot_ms_q25": 50.9127735,
    "steady_tpot_ms_q75": 51.0554525,
    "steady_tpot_p95_ms": 51.880345350000006,
    "steady_tpot_p95_ms_q25": 51.671559775000006,
    "steady_tpot_p95_ms_q75": 54.092308575000004,
    "decode_tokens_per_s": 17.818719397930845,
    "decode_tokens_per_s_q25": 17.223284756227333,
    "decode_tokens_per_s_q75": 17.871784564224235,
    "decode_ms": 841.8113370000001,
    "decode_ms_q25": 839.3192180000001,
    "decode_ms_q75": 871.956162,
    "end_to_end_ms": 3054.0277400000004,
    "end_to_end_ms_q25": 3047.6601100000003,
    "end_to_end_ms_q75": 3098.945381,
    "output_tokens_per_s_e2e": 5.238983192732885,
    "output_tokens_per_s_e2e_q25": 5.164131723502363,
    "output_tokens_per_s_e2e_q75": 5.24995218264981,
    "all_tokens_per_s_e2e": 1341.1796973396185,
    "all_tokens_per_s_e2e_q25": 1322.017721216605,
    "all_tokens_per_s_e2e_q75": 1343.9877587583514,
    "energy_net_psys_j": 70.63349411528053,
    "energy_net_psys_j_q25": 69.99021321995122,
    "energy_net_psys_j_q75": 70.91537397998445,
    "prefill_energy_net_psys_j": 51.578054242284516,
    "prefill_energy_net_psys_j_q25": 51.06518641593698,
    "prefill_energy_net_psys_j_q75": 51.911024644162225,
    "decode_energy_net_psys_j": 18.953258798648438,
    "decode_energy_net_psys_j_q25": 18.873936266840467,
    "decode_energy_net_psys_j_q75": 19.004349335822223,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.8206942904192479,
    "energy_reduction_vs_gpu": 0.13327813418458756
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 4096,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2212.2164030000004,
    "ttft_ms_q25": 2208.340892,
    "ttft_ms_q75": 2226.989219,
    "prefill_tokens_per_s": 1844.3041984803508,
    "prefill_tokens_per_s_q25": 1832.1505594469477,
    "prefill_tokens_per_s_q75": 1847.5465361782622,
    "handoff_plus_first_decode_ms": 72.596193,
    "handoff_plus_first_decode_ms_q25": 68.940962,
    "handoff_plus_first_decode_ms_q75": 74.4178035,
    "steady_tpot_ms": 44.69485,
    "steady_tpot_ms_q25": 44.68039125,
    "steady_tpot_ms_q75": 44.700368,
    "steady_tpot_p95_ms": 47.106267450000004,
    "steady_tpot_p95_ms_q25": 46.839246125,
    "steady_tpot_p95_ms_q75": 47.156936875,
    "decode_tokens_per_s": 21.282415866569966,
    "decode_tokens_per_s_q25": 21.25051559404212,
    "decode_tokens_per_s_q75": 21.39729160432636,
    "decode_ms": 704.8072030000001,
    "decode_ms_q25": 701.0435074999999,
    "decode_ms_q75": 705.8668170000001,
    "end_to_end_ms": 2917.0236060000007,
    "end_to_end_ms_q25": 2909.3843995,
    "end_to_end_ms_q75": 2932.856036,
    "output_tokens_per_s_e2e": 5.485043030536242,
    "output_tokens_per_s_e2e_q25": 5.455592120130046,
    "output_tokens_per_s_e2e_q75": 5.499483091115212,
    "all_tokens_per_s_e2e": 1404.171015817278,
    "all_tokens_per_s_e2e_q25": 1396.6315827532917,
    "all_tokens_per_s_e2e_q75": 1407.8676713254943,
    "energy_net_psys_j": 73.8895635695215,
    "energy_net_psys_j_q25": 73.01486646323396,
    "energy_net_psys_j_q75": 73.91778531588417,
    "prefill_energy_net_psys_j": 51.578054242284516,
    "prefill_energy_net_psys_j_q25": 51.06518641593698,
    "prefill_energy_net_psys_j_q75": 51.911024644162225,
    "decode_energy_net_psys_j": 21.645568523481565,
    "decode_energy_net_psys_j_q25": 21.616709645419277,
    "decode_energy_net_psys_j_q75": 22.006760671721942,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.8592399197060181,
    "energy_reduction_vs_gpu": 0.09332390810597446
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 8192,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 4599.011804,
    "ttft_ms_q25": 4485.462568,
    "ttft_ms_q75": 4602.603225499999,
    "prefill_tokens_per_s": 1777.773214865182,
    "prefill_tokens_per_s_q25": 1776.3870960761835,
    "prefill_tokens_per_s_q75": 1823.9463146042976,
    "handoff_plus_first_decode_ms": 79.860898,
    "handoff_plus_first_decode_ms_q25": 79.601507,
    "handoff_plus_first_decode_ms_q75": 82.1110245,
    "steady_tpot_ms": 48.571490000000004,
    "steady_tpot_ms_q25": 48.4244975,
    "steady_tpot_ms_q75": 48.717301250000006,
    "steady_tpot_p95_ms": 50.668234749999996,
    "steady_tpot_p95_ms_q25": 50.598343475,
    "steady_tpot_p95_ms_q75": 50.681439375,
    "decode_tokens_per_s": 19.545877697122307,
    "decode_tokens_per_s_q25": 19.53909174218755,
    "decode_tokens_per_s_q75": 19.56456280307617,
    "decode_ms": 767.4252459999998,
    "decode_ms_q25": 766.693017,
    "decode_ms_q75": 767.6918664999998,
    "end_to_end_ms": 5366.970291,
    "end_to_end_ms_q25": 5253.1544345,
    "end_to_end_ms_q75": 5369.562862999999,
    "output_tokens_per_s_e2e": 2.981197795491952,
    "output_tokens_per_s_e2e_q25": 2.9797590862020646,
    "output_tokens_per_s_e2e_q75": 3.047219439199628,
    "all_tokens_per_s_e2e": 1526.3732712918795,
    "all_tokens_per_s_e2e_q25": 1525.636652135457,
    "all_tokens_per_s_e2e_q75": 1560.1763528702095,
    "energy_net_psys_j": 177.1343408345819,
    "energy_net_psys_j_q25": 173.98962274979044,
    "energy_net_psys_j_q75": 178.29947360100846,
    "prefill_energy_net_psys_j": 152.23106387208279,
    "prefill_energy_net_psys_j_q25": 149.36397237921358,
    "prefill_energy_net_psys_j_q75": 153.26038140984252,
    "decode_energy_net_psys_j": 24.903276962499113,
    "decode_energy_net_psys_j_q25": 24.62565037057687,
    "decode_energy_net_psys_j_q75": 25.039092191165942,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 8192,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 4599.011804,
    "ttft_ms_q25": 4485.462568,
    "ttft_ms_q75": 4602.603225499999,
    "prefill_tokens_per_s": 1777.773214865182,
    "prefill_tokens_per_s_q25": 1776.3870960761835,
    "prefill_tokens_per_s_q75": 1823.9463146042976,
    "handoff_plus_first_decode_ms": 154.256099,
    "handoff_plus_first_decode_ms_q25": 154.0629455,
    "handoff_plus_first_decode_ms_q75": 176.731884,
    "steady_tpot_ms": 55.3727985,
    "steady_tpot_ms_q25": 55.34510075,
    "steady_tpot_ms_q75": 55.6846465,
    "steady_tpot_p95_ms": 57.49918405,
    "steady_tpot_p95_ms_q25": 57.098385775,
    "steady_tpot_p95_ms_q75": 58.010706225,
    "decode_tokens_per_s": 16.07177991201729,
    "decode_tokens_per_s_q25": 15.622681807816047,
    "decode_tokens_per_s_q75": 16.075234784006504,
    "decode_ms": 933.31293,
    "decode_ms_q25": 933.1123865000004,
    "decode_ms_q75": 960.9365339999995,
    "end_to_end_ms": 5531.9236470000005,
    "end_to_end_ms_q25": 5446.1985585,
    "end_to_end_ms_q75": 5535.715612,
    "output_tokens_per_s_e2e": 2.8923031156941055,
    "output_tokens_per_s_e2e_q25": 2.890323244373117,
    "output_tokens_per_s_e2e_q75": 2.9385570412007462,
    "all_tokens_per_s_e2e": 1480.859195235382,
    "all_tokens_per_s_e2e_q25": 1479.8455011190358,
    "all_tokens_per_s_e2e_q75": 1504.541205094782,
    "energy_net_psys_j": 172.29413562317183,
    "energy_net_psys_j_q25": 169.7769538779724,
    "energy_net_psys_j_q75": 172.7336527943006,
    "prefill_energy_net_psys_j": 152.23106387208279,
    "prefill_energy_net_psys_j_q25": 149.36397237921358,
    "prefill_energy_net_psys_j_q75": 153.26038140984252,
    "decode_energy_net_psys_j": 20.063071751089048,
    "decode_energy_net_psys_j_q25": 19.473271384458094,
    "decode_energy_net_psys_j_q75": 20.412981498758846,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9701815559060624,
    "energy_reduction_vs_gpu": 0.027325052774098357
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 8192,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 5483.789028,
    "ttft_ms_q25": 5470.867973,
    "ttft_ms_q75": 5499.028844500001,
    "prefill_tokens_per_s": 1490.9399246130151,
    "prefill_tokens_per_s_q25": 1486.819405236742,
    "prefill_tokens_per_s_q75": 1494.469551605921,
    "handoff_plus_first_decode_ms": 154.256099,
    "handoff_plus_first_decode_ms_q25": 154.0629455,
    "handoff_plus_first_decode_ms_q75": 176.731884,
    "steady_tpot_ms": 55.3727985,
    "steady_tpot_ms_q25": 55.34510075,
    "steady_tpot_ms_q75": 55.6846465,
    "steady_tpot_p95_ms": 57.49918405,
    "steady_tpot_p95_ms_q25": 57.098385775,
    "steady_tpot_p95_ms_q75": 58.010706225,
    "decode_tokens_per_s": 16.07177991201729,
    "decode_tokens_per_s_q25": 15.622681807816047,
    "decode_tokens_per_s_q75": 16.075234784006504,
    "decode_ms": 933.31293,
    "decode_ms_q25": 933.1123865000004,
    "decode_ms_q75": 960.9365339999995,
    "end_to_end_ms": 6446.5070559999995,
    "end_to_end_ms_q25": 6431.6039635,
    "end_to_end_ms_q75": 6447.0443235,
    "output_tokens_per_s_e2e": 2.481964242187281,
    "output_tokens_per_s_e2e_q25": 2.48175742375394,
    "output_tokens_per_s_e2e_q75": 2.4877287220074606,
    "all_tokens_per_s_e2e": 1270.7656919998878,
    "all_tokens_per_s_e2e_q25": 1270.6598009620172,
    "all_tokens_per_s_e2e_q75": 1273.7171056678199,
    "energy_net_psys_j": 145.28672479830834,
    "energy_net_psys_j_q25": 145.10051649871804,
    "energy_net_psys_j_q75": 149.32261254674575,
    "prefill_energy_net_psys_j": 126.03083718130063,
    "prefill_energy_net_psys_j_q25": 125.62724511425996,
    "prefill_energy_net_psys_j_q75": 129.3132231150276,
    "decode_energy_net_psys_j": 20.063071751089048,
    "decode_energy_net_psys_j_q25": 19.473271384458094,
    "decode_energy_net_psys_j_q75": 20.412981498758846,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.8325392719464666,
    "energy_reduction_vs_gpu": 0.17979357298094256
  },
  {
    "precision": "FP16",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 8192,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 5483.789028,
    "ttft_ms_q25": 5470.867973,
    "ttft_ms_q75": 5499.028844500001,
    "prefill_tokens_per_s": 1490.9399246130151,
    "prefill_tokens_per_s_q25": 1486.819405236742,
    "prefill_tokens_per_s_q75": 1494.469551605921,
    "handoff_plus_first_decode_ms": 79.860898,
    "handoff_plus_first_decode_ms_q25": 79.601507,
    "handoff_plus_first_decode_ms_q75": 82.1110245,
    "steady_tpot_ms": 48.571490000000004,
    "steady_tpot_ms_q25": 48.4244975,
    "steady_tpot_ms_q75": 48.717301250000006,
    "steady_tpot_p95_ms": 50.668234749999996,
    "steady_tpot_p95_ms_q25": 50.598343475,
    "steady_tpot_p95_ms_q75": 50.681439375,
    "decode_tokens_per_s": 19.545877697122307,
    "decode_tokens_per_s_q25": 19.53909174218755,
    "decode_tokens_per_s_q75": 19.56456280307617,
    "decode_ms": 767.4252459999998,
    "decode_ms_q25": 766.693017,
    "decode_ms_q75": 767.6918664999998,
    "end_to_end_ms": 6251.747515,
    "end_to_end_ms_q25": 6238.5598395,
    "end_to_end_ms_q75": 6265.988482000001,
    "output_tokens_per_s_e2e": 2.559284417934463,
    "output_tokens_per_s_e2e_q25": 2.553481017718692,
    "output_tokens_per_s_e2e_q75": 2.5647059431905914,
    "all_tokens_per_s_e2e": 1310.353621982445,
    "all_tokens_per_s_e2e_q25": 1307.3822810719703,
    "all_tokens_per_s_e2e_q75": 1313.1294429135828,
    "energy_net_psys_j": 151.2057446011334,
    "energy_net_psys_j_q25": 150.6663373054259,
    "energy_net_psys_j_q75": 154.0746887142713,
    "prefill_energy_net_psys_j": 126.03083718130063,
    "prefill_energy_net_psys_j_q25": 125.62724511425996,
    "prefill_energy_net_psys_j_q75": 129.3132231150276,
    "decode_energy_net_psys_j": 24.903276962499113,
    "decode_energy_net_psys_j_q25": 24.62565037057687,
    "decode_energy_net_psys_j_q75": 25.039092191165942,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.8584752148295931,
    "energy_reduction_vs_gpu": 0.14637814503547952
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 607.063275,
    "ttft_ms_q25": 594.2362655,
    "ttft_ms_q75": 618.426575,
    "prefill_tokens_per_s": 1660.4529404286563,
    "prefill_tokens_per_s_q25": 1630.4933846355866,
    "prefill_tokens_per_s_q75": 1697.0857346633275,
    "handoff_plus_first_decode_ms": 110.586364,
    "handoff_plus_first_decode_ms_q25": 107.9654845,
    "handoff_plus_first_decode_ms_q75": 111.99283550000001,
    "steady_tpot_ms": 90.976238,
    "steady_tpot_ms_q25": 90.66058025000001,
    "steady_tpot_ms_q75": 91.04845574999999,
    "steady_tpot_p95_ms": 95.74199429999999,
    "steady_tpot_p95_ms_q25": 95.41095737499998,
    "steady_tpot_p95_ms_q75": 96.798083775,
    "decode_tokens_per_s": 10.791038897709017,
    "decode_tokens_per_s_q25": 10.707293359903568,
    "decode_tokens_per_s_q75": 10.807628278604277,
    "decode_ms": 1390.0422509999999,
    "decode_ms_q25": 1387.9118485,
    "decode_ms_q75": 1400.9999685,
    "end_to_end_ms": 1997.1055259999998,
    "end_to_end_ms_q25": 1995.236234,
    "end_to_end_ms_q75": 2006.3384234999999,
    "output_tokens_per_s_e2e": 8.011594676244464,
    "output_tokens_per_s_e2e_q25": 7.974895288985564,
    "output_tokens_per_s_e2e_q75": 8.019107597985064,
    "all_tokens_per_s_e2e": 512.7420592796457,
    "all_tokens_per_s_e2e_q25": 510.3932984950761,
    "all_tokens_per_s_e2e_q75": 513.2228862710441,
    "energy_net_psys_j": 65.03463419199078,
    "energy_net_psys_j_q25": 63.088387584242376,
    "energy_net_psys_j_q75": 65.397238369353,
    "prefill_energy_net_psys_j": 21.91163438025128,
    "prefill_energy_net_psys_j_q25": 20.994881327407033,
    "prefill_energy_net_psys_j_q75": 22.062885108962686,
    "decode_energy_net_psys_j": 43.1229998117395,
    "decode_energy_net_psys_j_q25": 42.09350625683534,
    "decode_energy_net_psys_j_q75": 43.334353260390316,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 607.063275,
    "ttft_ms_q25": 594.2362655,
    "ttft_ms_q75": 618.426575,
    "prefill_tokens_per_s": 1660.4529404286563,
    "prefill_tokens_per_s_q25": 1630.4933846355866,
    "prefill_tokens_per_s_q75": 1697.0857346633275,
    "handoff_plus_first_decode_ms": 103.587636,
    "handoff_plus_first_decode_ms_q25": 103.343637,
    "handoff_plus_first_decode_ms_q75": 142.3081965,
    "steady_tpot_ms": 97.750102,
    "steady_tpot_ms_q25": 97.70124425,
    "steady_tpot_ms_q75": 97.83212674999999,
    "steady_tpot_p95_ms": 98.4875902,
    "steady_tpot_p95_ms_q25": 98.29367895,
    "steady_tpot_p95_ms_q75": 99.01323395,
    "decode_tokens_per_s": 10.186762902220105,
    "decode_tokens_per_s_q25": 9.914916723458596,
    "decode_tokens_per_s_q75": 10.18754351163878,
    "decode_ms": 1472.499178,
    "decode_ms_q25": 1472.386358,
    "decode_ms_q75": 1514.0101544999998,
    "end_to_end_ms": 2102.289053,
    "end_to_end_ms_q25": 2090.812933,
    "end_to_end_ms_q75": 2119.60972,
    "output_tokens_per_s_e2e": 7.610751707605454,
    "output_tokens_per_s_e2e_q25": 7.5490635511970545,
    "output_tokens_per_s_e2e_q75": 7.652756398399472,
    "all_tokens_per_s_e2e": 487.08810928674905,
    "all_tokens_per_s_e2e_q25": 483.1400672766115,
    "all_tokens_per_s_e2e_q75": 489.7764094975662,
    "energy_net_psys_j": 51.82984807263372,
    "energy_net_psys_j_q25": 51.07475836656299,
    "energy_net_psys_j_q75": 52.30109249696879,
    "prefill_energy_net_psys_j": 21.91163438025128,
    "prefill_energy_net_psys_j_q25": 20.994881327407033,
    "prefill_energy_net_psys_j_q75": 22.062885108962686,
    "decode_energy_net_psys_j": 30.241540385929476,
    "decode_energy_net_psys_j_q25": 30.07987703915596,
    "decode_energy_net_psys_j_q75": 30.399870734779622,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.949967143267049,
    "energy_reduction_vs_gpu": 0.203042367861635
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 738.013839,
    "ttft_ms_q25": 737.1676259999999,
    "ttft_ms_q75": 738.732179,
    "prefill_tokens_per_s": 1365.8280464846407,
    "prefill_tokens_per_s_q25": 1364.5012113635582,
    "prefill_tokens_per_s_q75": 1367.3977162352335,
    "handoff_plus_first_decode_ms": 103.587636,
    "handoff_plus_first_decode_ms_q25": 103.343637,
    "handoff_plus_first_decode_ms_q75": 142.3081965,
    "steady_tpot_ms": 97.750102,
    "steady_tpot_ms_q25": 97.70124425,
    "steady_tpot_ms_q75": 97.83212674999999,
    "steady_tpot_p95_ms": 98.4875902,
    "steady_tpot_p95_ms_q25": 98.29367895,
    "steady_tpot_p95_ms_q75": 99.01323395,
    "decode_tokens_per_s": 10.186762902220105,
    "decode_tokens_per_s_q25": 9.914916723458596,
    "decode_tokens_per_s_q75": 10.18754351163878,
    "decode_ms": 1472.499178,
    "decode_ms_q25": 1472.386358,
    "decode_ms_q75": 1514.0101544999998,
    "end_to_end_ms": 2211.949697,
    "end_to_end_ms_q25": 2210.272324,
    "end_to_end_ms_q75": 2252.7423335,
    "output_tokens_per_s_e2e": 7.233437551360373,
    "output_tokens_per_s_e2e_q25": 7.104784187649903,
    "output_tokens_per_s_e2e_q75": 7.238931167489643,
    "all_tokens_per_s_e2e": 462.94000328706386,
    "all_tokens_per_s_e2e_q25": 454.7061880095938,
    "all_tokens_per_s_e2e_q75": 463.2915947193371,
    "energy_net_psys_j": 51.049921114854904,
    "energy_net_psys_j_q25": 50.60905824675217,
    "energy_net_psys_j_q75": 51.057816044193146,
    "prefill_energy_net_psys_j": 20.491720031225135,
    "prefill_energy_net_psys_j_q25": 20.370850858746067,
    "prefill_energy_net_psys_j_q75": 20.65794530941352,
    "decode_energy_net_psys_j": 30.241540385929476,
    "decode_energy_net_psys_j_q25": 30.07987703915596,
    "decode_energy_net_psys_j_q75": 30.399870734779622,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.9028711316123569,
    "energy_reduction_vs_gpu": 0.21503485413404744
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 738.013839,
    "ttft_ms_q25": 737.1676259999999,
    "ttft_ms_q75": 738.732179,
    "prefill_tokens_per_s": 1365.8280464846407,
    "prefill_tokens_per_s_q25": 1364.5012113635582,
    "prefill_tokens_per_s_q75": 1367.3977162352335,
    "handoff_plus_first_decode_ms": 110.586364,
    "handoff_plus_first_decode_ms_q25": 107.9654845,
    "handoff_plus_first_decode_ms_q75": 111.99283550000001,
    "steady_tpot_ms": 90.976238,
    "steady_tpot_ms_q25": 90.66058025000001,
    "steady_tpot_ms_q75": 91.04845574999999,
    "steady_tpot_p95_ms": 95.74199429999999,
    "steady_tpot_p95_ms_q25": 95.41095737499998,
    "steady_tpot_p95_ms_q75": 96.798083775,
    "decode_tokens_per_s": 10.791038897709017,
    "decode_tokens_per_s_q25": 10.707293359903568,
    "decode_tokens_per_s_q75": 10.807628278604277,
    "decode_ms": 1390.0422509999999,
    "decode_ms_q25": 1387.9118485,
    "decode_ms_q75": 1400.9999685,
    "end_to_end_ms": 2126.363664,
    "end_to_end_ms_q25": 2125.7978144999997,
    "end_to_end_ms_q75": 2138.1675944999997,
    "output_tokens_per_s_e2e": 7.524583057397467,
    "output_tokens_per_s_e2e_q25": 7.483271042601831,
    "output_tokens_per_s_e2e_q75": 7.526586500612541,
    "all_tokens_per_s_e2e": 481.5733156734379,
    "all_tokens_per_s_e2e_q25": 478.92934672651717,
    "all_tokens_per_s_e2e_q75": 481.7015360392026,
    "energy_net_psys_j": 63.372981498006496,
    "energy_net_psys_j_q25": 62.6305823937698,
    "energy_net_psys_j_q75": 63.70520411913638,
    "prefill_energy_net_psys_j": 20.491720031225135,
    "prefill_energy_net_psys_j_q25": 20.370850858746067,
    "prefill_energy_net_psys_j_q75": 20.65794530941352,
    "decode_energy_net_psys_j": 43.1229998117395,
    "decode_energy_net_psys_j_q25": 42.09350625683534,
    "decode_energy_net_psys_j_q75": 43.334353260390316,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9392116502984035,
    "energy_reduction_vs_gpu": 0.025550273552379266
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1295.2486940000001,
    "ttft_ms_q25": 1268.754579,
    "ttft_ms_q75": 1301.2946175000002,
    "prefill_tokens_per_s": 1568.8106920415082,
    "prefill_tokens_per_s_q25": 1561.555574394189,
    "prefill_tokens_per_s_q75": 1602.2692557587686,
    "handoff_plus_first_decode_ms": 113.84433100000001,
    "handoff_plus_first_decode_ms_q25": 113.2854395,
    "handoff_plus_first_decode_ms_q75": 119.05460550000001,
    "steady_tpot_ms": 99.86533,
    "steady_tpot_ms_q25": 99.2980655,
    "steady_tpot_ms_q75": 99.88143000000001,
    "steady_tpot_p95_ms": 104.039077,
    "steady_tpot_p95_ms_q25": 103.99117665,
    "steady_tpot_p95_ms_q75": 104.07479005,
    "decode_tokens_per_s": 9.858521320098015,
    "decode_tokens_per_s_q25": 9.84239684144253,
    "decode_tokens_per_s_q75": 9.85915772427815,
    "decode_ms": 1521.5263540000003,
    "decode_ms_q25": 1521.4281465,
    "decode_ms_q75": 1524.0231115000001,
    "end_to_end_ms": 2816.7750480000004,
    "end_to_end_ms_q25": 2792.7776905,
    "end_to_end_ms_q75": 2822.722764,
    "output_tokens_per_s_e2e": 5.680254804642816,
    "output_tokens_per_s_e2e_q25": 5.668311192750048,
    "output_tokens_per_s_e2e_q75": 5.729486267180118,
    "all_tokens_per_s_e2e": 727.0726149942805,
    "all_tokens_per_s_e2e_q25": 725.5438326720061,
    "all_tokens_per_s_e2e_q75": 733.3742421990551,
    "energy_net_psys_j": 91.30615653603783,
    "energy_net_psys_j_q25": 89.48061312461171,
    "energy_net_psys_j_q75": 91.31300907224337,
    "prefill_energy_net_psys_j": 46.848244522061265,
    "prefill_energy_net_psys_j_q25": 45.756798212515726,
    "prefill_energy_net_psys_j_q75": 47.06509762227614,
    "decode_energy_net_psys_j": 44.037910885957885,
    "decode_energy_net_psys_j_q25": 43.51381434808664,
    "decode_energy_net_psys_j_q75": 44.24791144996723,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1295.2486940000001,
    "ttft_ms_q25": 1268.754579,
    "ttft_ms_q75": 1301.2946175000002,
    "prefill_tokens_per_s": 1568.8106920415082,
    "prefill_tokens_per_s_q25": 1561.555574394189,
    "prefill_tokens_per_s_q75": 1602.2692557587686,
    "handoff_plus_first_decode_ms": 134.185059,
    "handoff_plus_first_decode_ms_q25": 133.730439,
    "handoff_plus_first_decode_ms_q75": 157.284578,
    "steady_tpot_ms": 99.4194085,
    "steady_tpot_ms_q25": 99.33907400000001,
    "steady_tpot_ms_q75": 99.550228,
    "steady_tpot_p95_ms": 100.32548745,
    "steady_tpot_p95_ms_q25": 100.00277105,
    "steady_tpot_p95_ms_q75": 100.393350875,
    "decode_tokens_per_s": 9.815708810668962,
    "decode_tokens_per_s_q25": 9.662051614427256,
    "decode_tokens_per_s_q75": 9.82897138692405,
    "decode_ms": 1528.162692,
    "decode_ms_q25": 1526.103467,
    "decode_ms_q75": 1552.858049,
    "end_to_end_ms": 2823.411386,
    "end_to_end_ms_q25": 2821.612628,
    "end_to_end_ms_q75": 2827.3980844999996,
    "output_tokens_per_s_e2e": 5.66690354772126,
    "output_tokens_per_s_e2e_q25": 5.6589243299162515,
    "output_tokens_per_s_e2e_q75": 5.6705184629467436,
    "all_tokens_per_s_e2e": 725.3636541083213,
    "all_tokens_per_s_e2e_q25": 724.3423142292802,
    "all_tokens_per_s_e2e_q75": 725.8263632571832,
    "energy_net_psys_j": 74.87595080459188,
    "energy_net_psys_j_q25": 74.50002106049851,
    "energy_net_psys_j_q75": 76.09583793255571,
    "prefill_energy_net_psys_j": 46.848244522061265,
    "prefill_energy_net_psys_j_q25": 45.756798212515726,
    "prefill_energy_net_psys_j_q75": 47.06509762227614,
    "decode_energy_net_psys_j": 29.45873941343495,
    "decode_energy_net_psys_j_q25": 28.743222847982786,
    "decode_energy_net_psys_j_q75": 29.74625687573174,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9976495320402452,
    "energy_reduction_vs_gpu": 0.1799463076179434
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1491.060119,
    "ttft_ms_q25": 1488.32563,
    "ttft_ms_q75": 1502.124392,
    "prefill_tokens_per_s": 1362.7887796789769,
    "prefill_tokens_per_s_q25": 1352.824214487296,
    "prefill_tokens_per_s_q75": 1365.2972295643517,
    "handoff_plus_first_decode_ms": 134.185059,
    "handoff_plus_first_decode_ms_q25": 133.730439,
    "handoff_plus_first_decode_ms_q75": 157.284578,
    "steady_tpot_ms": 99.4194085,
    "steady_tpot_ms_q25": 99.33907400000001,
    "steady_tpot_ms_q75": 99.550228,
    "steady_tpot_p95_ms": 100.32548745,
    "steady_tpot_p95_ms_q25": 100.00277105,
    "steady_tpot_p95_ms_q75": 100.393350875,
    "decode_tokens_per_s": 9.815708810668962,
    "decode_tokens_per_s_q25": 9.662051614427256,
    "decode_tokens_per_s_q75": 9.82897138692405,
    "decode_ms": 1528.162692,
    "decode_ms_q25": 1526.103467,
    "decode_ms_q75": 1552.858049,
    "end_to_end_ms": 3041.351357,
    "end_to_end_ms_q25": 3028.2278589999996,
    "end_to_end_ms_q75": 3052.2479519999997,
    "output_tokens_per_s_e2e": 5.260819327294844,
    "output_tokens_per_s_e2e_q25": 5.242104892929072,
    "output_tokens_per_s_e2e_q75": 5.28371749056015,
    "all_tokens_per_s_e2e": 673.3848738937401,
    "all_tokens_per_s_e2e_q25": 670.9894262949213,
    "all_tokens_per_s_e2e_q75": 676.3158387916992,
    "energy_net_psys_j": 74.0294116716735,
    "energy_net_psys_j_q25": 72.20316558915493,
    "energy_net_psys_j_q75": 74.23449914619869,
    "prefill_energy_net_psys_j": 43.995637333644986,
    "prefill_energy_net_psys_j_q25": 43.172425278875366,
    "prefill_energy_net_psys_j_q75": 44.48824227046694,
    "decode_energy_net_psys_j": 29.45873941343495,
    "decode_energy_net_psys_j_q25": 28.743222847982786,
    "decode_energy_net_psys_j_q75": 29.74625687573174,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.9261590383225164,
    "energy_reduction_vs_gpu": 0.18921774302859118
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1491.060119,
    "ttft_ms_q25": 1488.32563,
    "ttft_ms_q75": 1502.124392,
    "prefill_tokens_per_s": 1362.7887796789769,
    "prefill_tokens_per_s_q25": 1352.824214487296,
    "prefill_tokens_per_s_q75": 1365.2972295643517,
    "handoff_plus_first_decode_ms": 113.84433100000001,
    "handoff_plus_first_decode_ms_q25": 113.2854395,
    "handoff_plus_first_decode_ms_q75": 119.05460550000001,
    "steady_tpot_ms": 99.86533,
    "steady_tpot_ms_q25": 99.2980655,
    "steady_tpot_ms_q75": 99.88143000000001,
    "steady_tpot_p95_ms": 104.039077,
    "steady_tpot_p95_ms_q25": 103.99117665,
    "steady_tpot_p95_ms_q75": 104.07479005,
    "decode_tokens_per_s": 9.858521320098015,
    "decode_tokens_per_s_q25": 9.84239684144253,
    "decode_tokens_per_s_q75": 9.85915772427815,
    "decode_ms": 1521.5263540000003,
    "decode_ms_q25": 1521.4281465,
    "decode_ms_q75": 1524.0231115000001,
    "end_to_end_ms": 3012.390058,
    "end_to_end_ms_q25": 3012.250534,
    "end_to_end_ms_q75": 3023.5525385,
    "output_tokens_per_s_e2e": 5.311397160373978,
    "output_tokens_per_s_e2e_q25": 5.291860443792936,
    "output_tokens_per_s_e2e_q75": 5.311643189612258,
    "all_tokens_per_s_e2e": 679.8588365278691,
    "all_tokens_per_s_e2e_q25": 677.3581368054959,
    "all_tokens_per_s_e2e_q75": 679.890328270369,
    "energy_net_psys_j": 87.9705650175043,
    "energy_net_psys_j_q25": 87.3888451277933,
    "energy_net_psys_j_q75": 88.00205661855358,
    "prefill_energy_net_psys_j": 43.995637333644986,
    "prefill_energy_net_psys_j_q25": 43.172425278875366,
    "prefill_energy_net_psys_j_q75": 44.48824227046694,
    "decode_energy_net_psys_j": 44.037910885957885,
    "decode_energy_net_psys_j_q25": 43.51381434808664,
    "decode_energy_net_psys_j_q75": 44.24791144996723,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9350631869599673,
    "energy_reduction_vs_gpu": 0.03653194532634829
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3240.103964,
    "ttft_ms_q25": 3180.682677,
    "ttft_ms_q75": 3252.5380395,
    "prefill_tokens_per_s": 1259.2188538799617,
    "prefill_tokens_per_s_q25": 1254.4233390183695,
    "prefill_tokens_per_s_q75": 1283.1913430639618,
    "handoff_plus_first_decode_ms": 115.762957,
    "handoff_plus_first_decode_ms_q25": 115.3336545,
    "handoff_plus_first_decode_ms_q75": 119.953497,
    "steady_tpot_ms": 92.687543,
    "steady_tpot_ms_q25": 92.273866,
    "steady_tpot_ms_q75": 92.77256249999999,
    "steady_tpot_p95_ms": 100.3072394,
    "steady_tpot_p95_ms_q25": 98.29813765,
    "steady_tpot_p95_ms_q75": 100.356425,
    "decode_tokens_per_s": 10.48851653020552,
    "decode_tokens_per_s_q25": 10.485455162324818,
    "decode_tokens_per_s_q75": 10.528354143165284,
    "decode_ms": 1430.1355159999998,
    "decode_ms_q25": 1424.74451,
    "decode_ms_q75": 1430.5531849999995,
    "end_to_end_ms": 4671.074817999999,
    "end_to_end_ms_q25": 4611.2358619999995,
    "end_to_end_ms_q75": 4677.7002185,
    "output_tokens_per_s_e2e": 3.4253358431220065,
    "output_tokens_per_s_e2e_q25": 3.42049112828928,
    "output_tokens_per_s_e2e_q75": 3.470370043618033,
    "all_tokens_per_s_e2e": 876.8859758392337,
    "all_tokens_per_s_e2e_q25": 875.6457288420556,
    "all_tokens_per_s_e2e_q75": 888.4147311662165,
    "energy_net_psys_j": 156.04669809514434,
    "energy_net_psys_j_q25": 154.98019578291135,
    "energy_net_psys_j_q75": 158.8095087872635,
    "prefill_energy_net_psys_j": 114.60094196753823,
    "prefill_energy_net_psys_j_q25": 113.44999763618469,
    "prefill_energy_net_psys_j_q75": 116.04362185376917,
    "decode_energy_net_psys_j": 41.61464016584719,
    "decode_energy_net_psys_j_q25": 41.53019814672665,
    "decode_energy_net_psys_j_q75": 42.85032895261487,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3240.103964,
    "ttft_ms_q25": 3180.682677,
    "ttft_ms_q75": 3252.5380395,
    "prefill_tokens_per_s": 1259.2188538799617,
    "prefill_tokens_per_s_q25": 1254.4233390183695,
    "prefill_tokens_per_s_q75": 1283.1913430639618,
    "handoff_plus_first_decode_ms": 205.440954,
    "handoff_plus_first_decode_ms_q25": 193.873883,
    "handoff_plus_first_decode_ms_q75": 213.41842350000002,
    "steady_tpot_ms": 102.098412,
    "steady_tpot_ms_q25": 101.98185075,
    "steady_tpot_ms_q75": 102.19718950000001,
    "steady_tpot_p95_ms": 103.19750105,
    "steady_tpot_p95_ms_q25": 102.936983825,
    "steady_tpot_p95_ms_q75": 103.788693575,
    "decode_tokens_per_s": 9.131076469404587,
    "decode_tokens_per_s_q25": 9.108240976025336,
    "decode_tokens_per_s_q75": 9.214446473562187,
    "decode_ms": 1642.7416909999997,
    "decode_ms_q25": 1628.0118475000002,
    "decode_ms_q75": 1646.8706005000001,
    "end_to_end_ms": 4853.3859680000005,
    "end_to_end_ms_q25": 4812.823434000001,
    "end_to_end_ms_q75": 4880.549887,
    "output_tokens_per_s_e2e": 3.2966675441626445,
    "output_tokens_per_s_e2e_q25": 3.2784206753191354,
    "output_tokens_per_s_e2e_q75": 3.3246880552262352,
    "all_tokens_per_s_e2e": 843.946891305637,
    "all_tokens_per_s_e2e_q25": 839.2756928816987,
    "all_tokens_per_s_e2e_q75": 851.1201421379162,
    "energy_net_psys_j": 145.01140496001767,
    "energy_net_psys_j_q25": 144.93403777621037,
    "energy_net_psys_j_q75": 146.639869330576,
    "prefill_energy_net_psys_j": 114.60094196753823,
    "prefill_energy_net_psys_j_q25": 113.44999763618469,
    "prefill_energy_net_psys_j_q75": 116.04362185376917,
    "decode_energy_net_psys_j": 30.782031961134198,
    "decode_energy_net_psys_j_q25": 30.596247476806816,
    "decode_energy_net_psys_j_q75": 31.669824624353048,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9624362968035017,
    "energy_reduction_vs_gpu": 0.07071788938717727
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3530.077925,
    "ttft_ms_q25": 3529.8153009999996,
    "ttft_ms_q75": 3551.0869350000003,
    "prefill_tokens_per_s": 1155.7818514728679,
    "prefill_tokens_per_s_q25": 1148.9842086185654,
    "prefill_tokens_per_s_q75": 1155.8678499210394,
    "handoff_plus_first_decode_ms": 205.440954,
    "handoff_plus_first_decode_ms_q25": 193.873883,
    "handoff_plus_first_decode_ms_q75": 213.41842350000002,
    "steady_tpot_ms": 102.098412,
    "steady_tpot_ms_q25": 101.98185075,
    "steady_tpot_ms_q75": 102.19718950000001,
    "steady_tpot_p95_ms": 103.19750105,
    "steady_tpot_p95_ms_q25": 102.936983825,
    "steady_tpot_p95_ms_q75": 103.788693575,
    "decode_tokens_per_s": 9.131076469404587,
    "decode_tokens_per_s_q25": 9.108240976025336,
    "decode_tokens_per_s_q75": 9.214446473562187,
    "decode_ms": 1642.7416909999997,
    "decode_ms_q25": 1628.0118475000002,
    "decode_ms_q75": 1646.8706005000001,
    "end_to_end_ms": 5172.819616,
    "end_to_end_ms_q25": 5157.8271485000005,
    "end_to_end_ms_q75": 5197.9575355,
    "output_tokens_per_s_e2e": 3.09309065224516,
    "output_tokens_per_s_e2e_q25": 3.078204103742986,
    "output_tokens_per_s_e2e_q75": 3.1021076757263275,
    "all_tokens_per_s_e2e": 791.831206974761,
    "all_tokens_per_s_e2e_q25": 788.0202505582045,
    "all_tokens_per_s_e2e_q75": 794.1395649859398,
    "energy_net_psys_j": 131.26549072083276,
    "energy_net_psys_j_q25": 131.13414926132015,
    "energy_net_psys_j_q75": 131.75081211012662,
    "prefill_energy_net_psys_j": 100.59234480932811,
    "prefill_energy_net_psys_j_q25": 99.65010912129449,
    "prefill_energy_net_psys_j_q75": 101.0232231738072,
    "decode_energy_net_psys_j": 30.782031961134198,
    "decode_energy_net_psys_j_q25": 30.596247476806816,
    "decode_energy_net_psys_j_q75": 31.669824624353048,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.9030036159683477,
    "energy_reduction_vs_gpu": 0.15880635525656595
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3530.077925,
    "ttft_ms_q25": 3529.8153009999996,
    "ttft_ms_q75": 3551.0869350000003,
    "prefill_tokens_per_s": 1155.7818514728679,
    "prefill_tokens_per_s_q25": 1148.9842086185654,
    "prefill_tokens_per_s_q75": 1155.8678499210394,
    "handoff_plus_first_decode_ms": 115.762957,
    "handoff_plus_first_decode_ms_q25": 115.3336545,
    "handoff_plus_first_decode_ms_q75": 119.953497,
    "steady_tpot_ms": 92.687543,
    "steady_tpot_ms_q25": 92.273866,
    "steady_tpot_ms_q75": 92.77256249999999,
    "steady_tpot_p95_ms": 100.3072394,
    "steady_tpot_p95_ms_q25": 98.29813765,
    "steady_tpot_p95_ms_q75": 100.356425,
    "decode_tokens_per_s": 10.48851653020552,
    "decode_tokens_per_s_q25": 10.485455162324818,
    "decode_tokens_per_s_q75": 10.528354143165284,
    "decode_ms": 1430.1355159999998,
    "decode_ms_q25": 1424.74451,
    "decode_ms_q75": 1430.5531849999995,
    "end_to_end_ms": 4960.523530999999,
    "end_to_end_ms_q25": 4954.97748,
    "end_to_end_ms_q75": 4981.377495999999,
    "output_tokens_per_s_e2e": 3.225466001725535,
    "output_tokens_per_s_e2e_q25": 3.21201925188337,
    "output_tokens_per_s_e2e_q75": 3.229080275243806,
    "all_tokens_per_s_e2e": 825.719296441737,
    "all_tokens_per_s_e2e_q25": 822.2769284821427,
    "all_tokens_per_s_e2e_q75": 826.6445504624144,
    "energy_net_psys_j": 142.0381009369342,
    "energy_net_psys_j_q25": 141.18030726802112,
    "energy_net_psys_j_q75": 143.78911010730152,
    "prefill_energy_net_psys_j": 100.59234480932811,
    "prefill_energy_net_psys_j_q25": 99.65010912129449,
    "prefill_energy_net_psys_j_q75": 101.0232231738072,
    "decode_energy_net_psys_j": 41.61464016584719,
    "decode_energy_net_psys_j_q25": 41.53019814672665,
    "decode_energy_net_psys_j_q75": 42.85032895261487,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9416495635609555,
    "energy_reduction_vs_gpu": 0.08977182682628027
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 7706.322755,
    "ttft_ms_q25": 7599.4729475,
    "ttft_ms_q75": 7721.44232,
    "prefill_tokens_per_s": 1060.9469989685113,
    "prefill_tokens_per_s_q25": 1058.8735899593285,
    "prefill_tokens_per_s_q75": 1076.0768113839117,
    "handoff_plus_first_decode_ms": 131.406359,
    "handoff_plus_first_decode_ms_q25": 130.64529349999998,
    "handoff_plus_first_decode_ms_q75": 135.692465,
    "steady_tpot_ms": 99.441204,
    "steady_tpot_ms_q25": 98.57963075,
    "steady_tpot_ms_q75": 99.513441,
    "steady_tpot_p95_ms": 107.64871504999999,
    "steady_tpot_p95_ms_q25": 107.460514375,
    "steady_tpot_p95_ms_q75": 107.72995429999999,
    "decode_tokens_per_s": 9.70446882113418,
    "decode_tokens_per_s_q25": 9.695234573694833,
    "decode_tokens_per_s_q75": 9.779850182263372,
    "decode_ms": 1545.6796530000001,
    "decode_ms_q25": 1533.8569539999999,
    "decode_ms_q75": 1547.1532425,
    "end_to_end_ms": 9252.002408,
    "end_to_end_ms_q25": 9146.626189999999,
    "end_to_end_ms_q75": 9255.299274,
    "output_tokens_per_s_e2e": 1.7293553648629787,
    "output_tokens_per_s_e2e_q25": 1.7287395638475633,
    "output_tokens_per_s_e2e_q75": 1.7495110876326878,
    "all_tokens_per_s_e2e": 885.4299468098451,
    "all_tokens_per_s_e2e_q25": 885.1146566899524,
    "all_tokens_per_s_e2e_q75": 895.7496768679362,
    "energy_net_psys_j": 318.80381272461636,
    "energy_net_psys_j_q25": 316.7465021804097,
    "energy_net_psys_j_q75": 321.3544898898683,
    "prefill_energy_net_psys_j": 270.7895866055883,
    "prefill_energy_net_psys_j_q25": 269.2357969458508,
    "prefill_energy_net_psys_j_q75": 274.5292660496832,
    "decode_energy_net_psys_j": 47.00718435008983,
    "decode_energy_net_psys_j_q25": 46.32170295571602,
    "decode_energy_net_psys_j_q75": 47.510705234558955,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 7706.322755,
    "ttft_ms_q25": 7599.4729475,
    "ttft_ms_q75": 7721.44232,
    "prefill_tokens_per_s": 1060.9469989685113,
    "prefill_tokens_per_s_q25": 1058.8735899593285,
    "prefill_tokens_per_s_q75": 1076.0768113839117,
    "handoff_plus_first_decode_ms": 250.5723,
    "handoff_plus_first_decode_ms_q25": 249.529976,
    "handoff_plus_first_decode_ms_q75": 269.1960525,
    "steady_tpot_ms": 108.125211,
    "steady_tpot_ms_q25": 108.08177275,
    "steady_tpot_ms_q75": 108.156571,
    "steady_tpot_p95_ms": 110.94088165,
    "steady_tpot_p95_ms_q25": 110.91701512499999,
    "steady_tpot_p95_ms_q75": 110.98606692499999,
    "decode_tokens_per_s": 8.465157969907148,
    "decode_tokens_per_s_q25": 8.37940825708966,
    "decode_tokens_per_s_q75": 8.473290083584455,
    "decode_ms": 1771.969294999999,
    "decode_ms_q25": 1770.270304499999,
    "decode_ms_q75": 1790.2900224999994,
    "end_to_end_ms": 9474.894068999998,
    "end_to_end_ms_q25": 9388.063979499999,
    "end_to_end_ms_q75": 9491.712624499998,
    "output_tokens_per_s_e2e": 1.6886732330178629,
    "output_tokens_per_s_e2e_q25": 1.6856863316812385,
    "output_tokens_per_s_e2e_q75": 1.7044375552242785,
    "all_tokens_per_s_e2e": 864.6006953051458,
    "all_tokens_per_s_e2e_q25": 863.0714018207941,
    "all_tokens_per_s_e2e_q75": 872.6720282748306,
    "energy_net_psys_j": 301.62856917031655,
    "energy_net_psys_j_q25": 301.57297758855043,
    "energy_net_psys_j_q75": 307.59249950694516,
    "prefill_energy_net_psys_j": 270.7895866055883,
    "prefill_energy_net_psys_j_q25": 269.2357969458508,
    "prefill_energy_net_psys_j_q75": 274.5292660496832,
    "decode_energy_net_psys_j": 33.946561884203355,
    "decode_energy_net_psys_j_q25": 32.33718064269972,
    "decode_energy_net_psys_j_q75": 34.61702311699953,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.9764755511379007,
    "energy_reduction_vs_gpu": 0.0538740217926309
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 8873.004211,
    "ttft_ms_q25": 8847.967757,
    "ttft_ms_q75": 8873.702487,
    "prefill_tokens_per_s": 921.4466493619024,
    "prefill_tokens_per_s_q25": 921.3741459785514,
    "prefill_tokens_per_s_q75": 924.0613990247903,
    "handoff_plus_first_decode_ms": 250.5723,
    "handoff_plus_first_decode_ms_q25": 249.529976,
    "handoff_plus_first_decode_ms_q75": 269.1960525,
    "steady_tpot_ms": 108.125211,
    "steady_tpot_ms_q25": 108.08177275,
    "steady_tpot_ms_q75": 108.156571,
    "steady_tpot_p95_ms": 110.94088165,
    "steady_tpot_p95_ms_q25": 110.91701512499999,
    "steady_tpot_p95_ms_q75": 110.98606692499999,
    "decode_tokens_per_s": 8.465157969907148,
    "decode_tokens_per_s_q25": 8.37940825708966,
    "decode_tokens_per_s_q75": 8.473290083584455,
    "decode_ms": 1771.969294999999,
    "decode_ms_q25": 1770.270304499999,
    "decode_ms_q75": 1790.2900224999994,
    "end_to_end_ms": 10646.370058,
    "end_to_end_ms_q25": 10618.9363375,
    "end_to_end_ms_q75": 10663.9925095,
    "output_tokens_per_s_e2e": 1.502859651959695,
    "output_tokens_per_s_e2e_q25": 1.5003802449201546,
    "output_tokens_per_s_e2e_q75": 1.5067523037517474,
    "all_tokens_per_s_e2e": 769.4641418033639,
    "all_tokens_per_s_e2e_q25": 768.1946853991192,
    "all_tokens_per_s_e2e_q75": 771.4571795208947,
    "energy_net_psys_j": 276.93145829420837,
    "energy_net_psys_j_q25": 275.7152032005781,
    "energy_net_psys_j_q75": 278.7892887981194,
    "prefill_energy_net_psys_j": 243.77114870575178,
    "prefill_energy_net_psys_j_q25": 242.70756132508222,
    "prefill_energy_net_psys_j_q75": 245.23585306178947,
    "decode_energy_net_psys_j": 33.946561884203355,
    "decode_energy_net_psys_j_q25": 32.33718064269972,
    "decode_energy_net_psys_j_q75": 34.61702311699953,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.8690288199260714,
    "energy_reduction_vs_gpu": 0.13134207546814203
  },
  {
    "precision": "FP16",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 8873.004211,
    "ttft_ms_q25": 8847.967757,
    "ttft_ms_q75": 8873.702487,
    "prefill_tokens_per_s": 921.4466493619024,
    "prefill_tokens_per_s_q25": 921.3741459785514,
    "prefill_tokens_per_s_q75": 924.0613990247903,
    "handoff_plus_first_decode_ms": 131.406359,
    "handoff_plus_first_decode_ms_q25": 130.64529349999998,
    "handoff_plus_first_decode_ms_q75": 135.692465,
    "steady_tpot_ms": 99.441204,
    "steady_tpot_ms_q25": 98.57963075,
    "steady_tpot_ms_q75": 99.513441,
    "steady_tpot_p95_ms": 107.64871504999999,
    "steady_tpot_p95_ms_q25": 107.460514375,
    "steady_tpot_p95_ms_q75": 107.72995429999999,
    "decode_tokens_per_s": 9.70446882113418,
    "decode_tokens_per_s_q25": 9.695234573694833,
    "decode_tokens_per_s_q75": 9.779850182263372,
    "decode_ms": 1545.6796530000001,
    "decode_ms_q25": 1533.8569539999999,
    "decode_ms_q75": 1547.1532425,
    "end_to_end_ms": 10396.435018,
    "end_to_end_ms_q25": 10382.522987,
    "end_to_end_ms_q75": 10409.033030499999,
    "output_tokens_per_s_e2e": 1.5389890834981603,
    "output_tokens_per_s_e2e_q25": 1.5371287025624625,
    "output_tokens_per_s_e2e_q75": 1.5410540142713511,
    "all_tokens_per_s_e2e": 787.9624107510581,
    "all_tokens_per_s_e2e_q25": 787.0098957119808,
    "all_tokens_per_s_e2e_q75": 789.0196553069318,
    "energy_net_psys_j": 291.78537482477986,
    "energy_net_psys_j_q25": 289.53278516526734,
    "energy_net_psys_j_q75": 292.74655829634844,
    "prefill_energy_net_psys_j": 243.77114870575178,
    "prefill_energy_net_psys_j_q25": 242.70756132508222,
    "prefill_energy_net_psys_j_q75": 245.23585306178947,
    "decode_energy_net_psys_j": 47.00718435008983,
    "decode_energy_net_psys_j_q25": 46.32170295571602,
    "decode_energy_net_psys_j_q75": 47.510705234558955,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.8899206691506683,
    "energy_reduction_vs_gpu": 0.0847494189888347
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 1024,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1122.3871139999999,
    "ttft_ms_q25": 1110.2382564999998,
    "ttft_ms_q75": 1125.5228419999999,
    "prefill_tokens_per_s": 898.0858630919743,
    "prefill_tokens_per_s_q25": 895.5907303191804,
    "prefill_tokens_per_s_q75": 908.0219549834039,
    "handoff_plus_first_decode_ms": 194.39052800000002,
    "handoff_plus_first_decode_ms_q25": 194.27981200000002,
    "handoff_plus_first_decode_ms_q75": 198.3308715,
    "steady_tpot_ms": 186.323086,
    "steady_tpot_ms_q25": 186.13135925,
    "steady_tpot_ms_q75": 186.55247525,
    "steady_tpot_p95_ms": 187.25145110000003,
    "steady_tpot_p95_ms_q25": 186.6876767,
    "steady_tpot_p95_ms_q75": 187.75809557500003,
    "decode_tokens_per_s": 5.34787218891652,
    "decode_tokens_per_s_q25": 5.340328811433944,
    "decode_tokens_per_s_q75": 5.348010202163604,
    "decode_ms": 2804.853869,
    "decode_ms_q25": 2804.7814875000004,
    "decode_ms_q75": 2808.821415,
    "end_to_end_ms": 3927.0962200000004,
    "end_to_end_ms_q25": 3915.019744,
    "end_to_end_ms_q75": 3934.2718755,
    "output_tokens_per_s_e2e": 4.0742571874136555,
    "output_tokens_per_s_e2e_q25": 4.066839743501149,
    "output_tokens_per_s_e2e_q75": 4.08686374245765,
    "all_tokens_per_s_e2e": 260.75245999447395,
    "all_tokens_per_s_e2e_q25": 260.27774358407356,
    "all_tokens_per_s_e2e_q75": 261.5592795172896,
    "energy_net_psys_j": 132.8522017517532,
    "energy_net_psys_j_q25": 131.33742346136717,
    "energy_net_psys_j_q75": 134.39923450420244,
    "prefill_energy_net_psys_j": 46.053012388909565,
    "prefill_energy_net_psys_j_q25": 45.59526162958316,
    "prefill_energy_net_psys_j_q75": 47.34791628297139,
    "decode_energy_net_psys_j": 86.79918936284363,
    "decode_energy_net_psys_j_q25": 85.74216183178399,
    "decode_energy_net_psys_j_q75": 87.05131822123104,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 1024,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1122.3871139999999,
    "ttft_ms_q25": 1110.2382564999998,
    "ttft_ms_q75": 1125.5228419999999,
    "prefill_tokens_per_s": 898.0858630919743,
    "prefill_tokens_per_s_q25": 895.5907303191804,
    "prefill_tokens_per_s_q75": 908.0219549834039,
    "handoff_plus_first_decode_ms": 213.00904400000002,
    "handoff_plus_first_decode_ms_q25": 212.45307200000002,
    "handoff_plus_first_decode_ms_q75": 246.69054150000002,
    "steady_tpot_ms": 205.83404,
    "steady_tpot_ms_q25": 205.77235474999998,
    "steady_tpot_ms_q75": 206.12990075,
    "steady_tpot_p95_ms": 206.5551994,
    "steady_tpot_p95_ms_q25": 206.36039689999998,
    "steady_tpot_p95_ms_q75": 207.17913145,
    "decode_tokens_per_s": 4.8466610673536,
    "decode_tokens_per_s_q25": 4.787817592065784,
    "decode_tokens_per_s_q75": 4.846979213151203,
    "decode_ms": 3094.914167,
    "decode_ms_q25": 3094.7110365,
    "decode_ms_q75": 3133.4247405,
    "end_to_end_ms": 4223.166476,
    "end_to_end_ms_q25": 4220.2338785,
    "end_to_end_ms_q75": 4246.5955945000005,
    "output_tokens_per_s_e2e": 3.788626399391791,
    "output_tokens_per_s_e2e_q25": 3.7678386562685864,
    "output_tokens_per_s_e2e_q75": 3.791260907947617,
    "all_tokens_per_s_e2e": 242.47208956107463,
    "all_tokens_per_s_e2e_q25": 241.14167400118953,
    "all_tokens_per_s_e2e_q75": 242.64069810864748,
    "energy_net_psys_j": 100.98954805108606,
    "energy_net_psys_j_q25": 98.63511666993068,
    "energy_net_psys_j_q75": 101.59815174569542,
    "prefill_energy_net_psys_j": 46.053012388909565,
    "prefill_energy_net_psys_j_q25": 45.59526162958316,
    "prefill_energy_net_psys_j_q75": 47.34791628297139,
    "decode_energy_net_psys_j": 52.346727874052846,
    "decode_energy_net_psys_j_q25": 51.74495114628571,
    "decode_energy_net_psys_j_q75": 54.25023546272402,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.9298937757527321,
    "energy_reduction_vs_gpu": 0.2398353454480604
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 1024,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1883.063757,
    "ttft_ms_q25": 1881.8492390000001,
    "ttft_ms_q75": 1888.3335175,
    "prefill_tokens_per_s": 535.2978603368659,
    "prefill_tokens_per_s_q25": 533.8081652392121,
    "prefill_tokens_per_s_q75": 535.6435568612792,
    "handoff_plus_first_decode_ms": 213.00904400000002,
    "handoff_plus_first_decode_ms_q25": 212.45307200000002,
    "handoff_plus_first_decode_ms_q75": 246.69054150000002,
    "steady_tpot_ms": 205.83404,
    "steady_tpot_ms_q25": 205.77235474999998,
    "steady_tpot_ms_q75": 206.12990075,
    "steady_tpot_p95_ms": 206.5551994,
    "steady_tpot_p95_ms_q25": 206.36039689999998,
    "steady_tpot_p95_ms_q75": 207.17913145,
    "decode_tokens_per_s": 4.8466610673536,
    "decode_tokens_per_s_q25": 4.787817592065784,
    "decode_tokens_per_s_q75": 4.846979213151203,
    "decode_ms": 3094.914167,
    "decode_ms_q25": 3094.7110365,
    "decode_ms_q75": 3133.4247405,
    "end_to_end_ms": 4988.111184,
    "end_to_end_ms_q25": 4983.044554,
    "end_to_end_ms_q75": 5020.340609500001,
    "output_tokens_per_s_e2e": 3.207626977386156,
    "output_tokens_per_s_e2e_q25": 3.1871661081684453,
    "output_tokens_per_s_e2e_q75": 3.2108917285201617,
    "all_tokens_per_s_e2e": 205.28812655271398,
    "all_tokens_per_s_e2e_q25": 203.9786309227805,
    "all_tokens_per_s_e2e_q75": 205.49707062529035,
    "energy_net_psys_j": 104.30911227310798,
    "energy_net_psys_j_q25": 102.39025491443726,
    "energy_net_psys_j_q75": 105.18933223341892,
    "prefill_energy_net_psys_j": 49.91580914233469,
    "prefill_energy_net_psys_j_q25": 49.62201613979134,
    "prefill_energy_net_psys_j_q75": 50.93909677069491,
    "decode_energy_net_psys_j": 52.346727874052846,
    "decode_energy_net_psys_j_q25": 51.74495114628571,
    "decode_energy_net_psys_j_q75": 54.25023546272402,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.7872912361289499,
    "energy_reduction_vs_gpu": 0.21484844889496568
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 1024,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1883.063757,
    "ttft_ms_q25": 1881.8492390000001,
    "ttft_ms_q75": 1888.3335175,
    "prefill_tokens_per_s": 535.2978603368659,
    "prefill_tokens_per_s_q25": 533.8081652392121,
    "prefill_tokens_per_s_q75": 535.6435568612792,
    "handoff_plus_first_decode_ms": 194.39052800000002,
    "handoff_plus_first_decode_ms_q25": 194.27981200000002,
    "handoff_plus_first_decode_ms_q75": 198.3308715,
    "steady_tpot_ms": 186.323086,
    "steady_tpot_ms_q25": 186.13135925,
    "steady_tpot_ms_q75": 186.55247525,
    "steady_tpot_p95_ms": 187.25145110000003,
    "steady_tpot_p95_ms_q25": 186.6876767,
    "steady_tpot_p95_ms_q75": 187.75809557500003,
    "decode_tokens_per_s": 5.34787218891652,
    "decode_tokens_per_s_q25": 5.340328811433944,
    "decode_tokens_per_s_q75": 5.348010202163604,
    "decode_ms": 2804.853869,
    "decode_ms_q25": 2804.7814875000004,
    "decode_ms_q75": 2808.821415,
    "end_to_end_ms": 4687.772863,
    "end_to_end_ms_q25": 4686.6307265,
    "end_to_end_ms_q75": 4697.0825509999995,
    "output_tokens_per_s_e2e": 3.413134652125742,
    "output_tokens_per_s_e2e_q25": 3.4063831496797774,
    "output_tokens_per_s_e2e_q75": 3.413966639140699,
    "all_tokens_per_s_e2e": 218.4406177360475,
    "all_tokens_per_s_e2e_q25": 218.00852157950575,
    "all_tokens_per_s_e2e_q75": 218.49386490500473,
    "energy_net_psys_j": 136.71499850517833,
    "energy_net_psys_j_q25": 135.36417797157532,
    "energy_net_psys_j_q75": 137.99041499192595,
    "prefill_energy_net_psys_j": 49.91580914233469,
    "prefill_energy_net_psys_j_q25": 49.62201613979134,
    "prefill_energy_net_psys_j_q75": 50.93909677069491,
    "decode_energy_net_psys_j": 86.79918936284363,
    "decode_energy_net_psys_j_q25": 85.74216183178399,
    "decode_energy_net_psys_j_q75": 87.05131822123104,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.8377317619196262,
    "energy_reduction_vs_gpu": -0.029075895638094984
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 2048,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2327.862173,
    "ttft_ms_q25": 2297.254624,
    "ttft_ms_q75": 2350.933028,
    "prefill_tokens_per_s": 872.9039131132442,
    "prefill_tokens_per_s_q25": 864.4209279957477,
    "prefill_tokens_per_s_q75": 884.6911249767875,
    "handoff_plus_first_decode_ms": 199.299217,
    "handoff_plus_first_decode_ms_q25": 197.535943,
    "handoff_plus_first_decode_ms_q75": 201.2887685,
    "steady_tpot_ms": 185.658828,
    "steady_tpot_ms_q25": 184.875764,
    "steady_tpot_ms_q75": 185.846622,
    "steady_tpot_p95_ms": 186.86304305000002,
    "steady_tpot_p95_ms_q25": 186.3178438,
    "steady_tpot_p95_ms_q75": 187.170955575,
    "decode_tokens_per_s": 5.359118528725976,
    "decode_tokens_per_s_q25": 5.353027761500256,
    "decode_tokens_per_s_q75": 5.372636984930659,
    "decode_ms": 2798.967763,
    "decode_ms_q25": 2791.9427665000003,
    "decode_ms_q75": 2802.1561045,
    "end_to_end_ms": 5126.829936,
    "end_to_end_ms_q25": 5089.1973904999995,
    "end_to_end_ms_q75": 5153.089132499999,
    "output_tokens_per_s_e2e": 3.1208368913604625,
    "output_tokens_per_s_e2e_q25": 3.1050143082922745,
    "output_tokens_per_s_e2e_q75": 3.1440861299460954,
    "all_tokens_per_s_e2e": 399.4671220941392,
    "all_tokens_per_s_e2e_q25": 397.44183146141114,
    "all_tokens_per_s_e2e_q75": 402.4430246331002,
    "energy_net_psys_j": 187.82896060698425,
    "energy_net_psys_j_q25": 186.08959630468215,
    "energy_net_psys_j_q75": 188.02331934598294,
    "prefill_energy_net_psys_j": 98.78065362709816,
    "prefill_energy_net_psys_j_q25": 98.32382270062134,
    "prefill_energy_net_psys_j_q75": 100.00239527306141,
    "decode_energy_net_psys_j": 86.60482368795958,
    "decode_energy_net_psys_j_q25": 86.54403195809756,
    "decode_energy_net_psys_j_q75": 88.02092407292153,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 2048,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2327.862173,
    "ttft_ms_q25": 2297.254624,
    "ttft_ms_q75": 2350.933028,
    "prefill_tokens_per_s": 872.9039131132442,
    "prefill_tokens_per_s_q25": 864.4209279957477,
    "prefill_tokens_per_s_q75": 884.6911249767875,
    "handoff_plus_first_decode_ms": 262.7809,
    "handoff_plus_first_decode_ms_q25": 259.80020149999996,
    "handoff_plus_first_decode_ms_q75": 278.6357845,
    "steady_tpot_ms": 208.02285700000002,
    "steady_tpot_ms_q25": 207.901767,
    "steady_tpot_ms_q75": 208.03618675,
    "steady_tpot_p95_ms": 209.2143584,
    "steady_tpot_p95_ms_q25": 208.609963225,
    "steady_tpot_p95_ms_q75": 209.366780475,
    "decode_tokens_per_s": 4.728065618031263,
    "decode_tokens_per_s_q25": 4.699739044758417,
    "decode_tokens_per_s_q75": 4.728628149741116,
    "decode_ms": 3172.544802,
    "decode_ms_q25": 3172.1674315,
    "decode_ms_q75": 3191.7825215000003,
    "end_to_end_ms": 5500.406975,
    "end_to_end_ms_q25": 5489.0371455,
    "end_to_end_ms_q75": 5523.1004594999995,
    "output_tokens_per_s_e2e": 2.9088756655138233,
    "output_tokens_per_s_e2e_q25": 2.896972496103935,
    "output_tokens_per_s_e2e_q75": 2.9149135313507983,
    "all_tokens_per_s_e2e": 372.3360851857694,
    "all_tokens_per_s_e2e_q25": 370.8124795013037,
    "all_tokens_per_s_e2e_q75": 373.1089320129022,
    "energy_net_psys_j": 157.52080229855835,
    "energy_net_psys_j_q25": 154.61754982706316,
    "energy_net_psys_j_q75": 158.27159191513144,
    "prefill_energy_net_psys_j": 98.78065362709816,
    "prefill_energy_net_psys_j_q25": 98.32382270062134,
    "prefill_energy_net_psys_j_q75": 100.00239527306141,
    "decode_energy_net_psys_j": 56.2966653795337,
    "decode_energy_net_psys_j_q25": 55.071985480478574,
    "decode_energy_net_psys_j_q75": 58.26919664207003,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.932081927628637,
    "energy_reduction_vs_gpu": 0.16136041114470667
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 2048,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3784.315485,
    "ttft_ms_q25": 3779.738423,
    "ttft_ms_q75": 3786.98877,
    "prefill_tokens_per_s": 536.9531182202691,
    "prefill_tokens_per_s_q25": 536.5743433596591,
    "prefill_tokens_per_s_q75": 537.6041281979026,
    "handoff_plus_first_decode_ms": 262.7809,
    "handoff_plus_first_decode_ms_q25": 259.80020149999996,
    "handoff_plus_first_decode_ms_q75": 278.6357845,
    "steady_tpot_ms": 208.02285700000002,
    "steady_tpot_ms_q25": 207.901767,
    "steady_tpot_ms_q75": 208.03618675,
    "steady_tpot_p95_ms": 209.2143584,
    "steady_tpot_p95_ms_q25": 208.609963225,
    "steady_tpot_p95_ms_q75": 209.366780475,
    "decode_tokens_per_s": 4.728065618031263,
    "decode_tokens_per_s_q25": 4.699739044758417,
    "decode_tokens_per_s_q75": 4.728628149741116,
    "decode_ms": 3172.544802,
    "decode_ms_q25": 3172.1674315,
    "decode_ms_q75": 3191.7825215000003,
    "end_to_end_ms": 6956.105546,
    "end_to_end_ms_q25": 6951.9058545,
    "end_to_end_ms_q75": 6978.393921,
    "output_tokens_per_s_e2e": 2.300137612086773,
    "output_tokens_per_s_e2e_q25": 2.292814564491957,
    "output_tokens_per_s_e2e_q75": 2.301527980094815,
    "all_tokens_per_s_e2e": 294.41761434710696,
    "all_tokens_per_s_e2e_q25": 293.4802642549705,
    "all_tokens_per_s_e2e_q75": 294.59558145213634,
    "energy_net_psys_j": 157.51794815752865,
    "energy_net_psys_j_q25": 157.41612531313348,
    "energy_net_psys_j_q75": 162.6573135078748,
    "prefill_energy_net_psys_j": 103.67064257610521,
    "prefill_energy_net_psys_j_q25": 102.34413983265489,
    "prefill_energy_net_psys_j_q75": 105.6127967648599,
    "decode_energy_net_psys_j": 56.2966653795337,
    "decode_energy_net_psys_j_q25": 55.071985480478574,
    "decode_energy_net_psys_j_q75": 58.26919664207003,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.7370258979103765,
    "energy_reduction_vs_gpu": 0.1613756065704839
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 2048,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3784.315485,
    "ttft_ms_q25": 3779.738423,
    "ttft_ms_q75": 3786.98877,
    "prefill_tokens_per_s": 536.9531182202691,
    "prefill_tokens_per_s_q25": 536.5743433596591,
    "prefill_tokens_per_s_q75": 537.6041281979026,
    "handoff_plus_first_decode_ms": 199.299217,
    "handoff_plus_first_decode_ms_q25": 197.535943,
    "handoff_plus_first_decode_ms_q75": 201.2887685,
    "steady_tpot_ms": 185.658828,
    "steady_tpot_ms_q25": 184.875764,
    "steady_tpot_ms_q75": 185.846622,
    "steady_tpot_p95_ms": 186.86304305000002,
    "steady_tpot_p95_ms_q25": 186.3178438,
    "steady_tpot_p95_ms_q75": 187.170955575,
    "decode_tokens_per_s": 5.359118528725976,
    "decode_tokens_per_s_q25": 5.353027761500256,
    "decode_tokens_per_s_q75": 5.372636984930659,
    "decode_ms": 2798.967763,
    "decode_ms_q25": 2791.9427665000003,
    "decode_ms_q75": 2802.1561045,
    "end_to_end_ms": 6574.579825,
    "end_to_end_ms_q25": 6574.3544745,
    "end_to_end_ms_q75": 6582.1198779999995,
    "output_tokens_per_s_e2e": 2.4336155961114976,
    "output_tokens_per_s_e2e_q25": 2.430830992540381,
    "output_tokens_per_s_e2e_q75": 2.4336990165130516,
    "all_tokens_per_s_e2e": 311.5027963022717,
    "all_tokens_per_s_e2e_q25": 311.14636704516874,
    "all_tokens_per_s_e2e_q75": 311.5134741136706,
    "energy_net_psys_j": 190.15388280434075,
    "energy_net_psys_j_q25": 188.88817179075244,
    "energy_net_psys_j_q75": 193.5729291079194,
    "prefill_energy_net_psys_j": 103.67064257610521,
    "prefill_energy_net_psys_j_q25": 102.34413983265489,
    "prefill_energy_net_psys_j_q75": 105.6127967648599,
    "decode_energy_net_psys_j": 86.60482368795958,
    "decode_energy_net_psys_j_q25": 86.54403195809756,
    "decode_energy_net_psys_j_q75": 88.02092407292153,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.779795830678807,
    "energy_reduction_vs_gpu": -0.012377868619638521
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 4096,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 5528.934798,
    "ttft_ms_q25": 5448.807011000001,
    "ttft_ms_q75": 5530.436725,
    "prefill_tokens_per_s": 737.9359947373357,
    "prefill_tokens_per_s_q25": 737.7356443602816,
    "prefill_tokens_per_s_q75": 748.9497237643218,
    "handoff_plus_first_decode_ms": 209.129247,
    "handoff_plus_first_decode_ms_q25": 207.3933045,
    "handoff_plus_first_decode_ms_q75": 210.9861555,
    "steady_tpot_ms": 188.47727250000003,
    "steady_tpot_ms_q25": 188.40073825000002,
    "steady_tpot_ms_q75": 188.802173,
    "steady_tpot_p95_ms": 190.72192264999998,
    "steady_tpot_p95_ms_q25": 190.219103625,
    "steady_tpot_p95_ms_q75": 191.32966362500002,
    "decode_tokens_per_s": 5.254028638369279,
    "decode_tokens_per_s_q25": 5.253332518057384,
    "decode_tokens_per_s_q75": 5.254524002663129,
    "decode_ms": 2854.9520820000007,
    "decode_ms_q25": 2854.68296,
    "decode_ms_q75": 2855.3304425,
    "end_to_end_ms": 8384.643601,
    "end_to_end_ms_q25": 8303.8683315,
    "end_to_end_ms_q75": 8385.7671675,
    "output_tokens_per_s_e2e": 1.9082504589809577,
    "output_tokens_per_s_e2e_q25": 1.9079948164124154,
    "output_tokens_per_s_e2e_q75": 1.9269951629427777,
    "all_tokens_per_s_e2e": 488.51211749912517,
    "all_tokens_per_s_e2e_q25": 488.44667300157835,
    "all_tokens_per_s_e2e_q75": 493.3107617133511,
    "energy_net_psys_j": 320.48796452606746,
    "energy_net_psys_j_q25": 317.6113090734332,
    "energy_net_psys_j_q75": 321.2715341199631,
    "prefill_energy_net_psys_j": 231.97102930961674,
    "prefill_energy_net_psys_j_q25": 228.48211675690027,
    "prefill_energy_net_psys_j_q75": 232.25588411777665,
    "decode_energy_net_psys_j": 89.51436478792222,
    "decode_energy_net_psys_j_q25": 89.01565000218648,
    "decode_energy_net_psys_j_q75": 89.62790710226865,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 4096,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 5528.934798,
    "ttft_ms_q25": 5448.807011000001,
    "ttft_ms_q75": 5530.436725,
    "prefill_tokens_per_s": 737.9359947373357,
    "prefill_tokens_per_s_q25": 737.7356443602816,
    "prefill_tokens_per_s_q75": 748.9497237643218,
    "handoff_plus_first_decode_ms": 330.0977,
    "handoff_plus_first_decode_ms_q25": 316.156475,
    "handoff_plus_first_decode_ms_q75": 332.4441515,
    "steady_tpot_ms": 214.40785449999998,
    "steady_tpot_ms_q25": 214.3305785,
    "steady_tpot_ms_q75": 214.41199475,
    "steady_tpot_p95_ms": 215.12930295,
    "steady_tpot_p95_ms_q25": 215.090931,
    "steady_tpot_p95_ms_q75": 215.477234925,
    "decode_tokens_per_s": 4.503460418454397,
    "decode_tokens_per_s_q25": 4.49964235404687,
    "decode_tokens_per_s_q75": 4.518838452798773,
    "decode_ms": 3330.77203,
    "decode_ms_q25": 3319.475541,
    "decode_ms_q75": 3333.6006774999996,
    "end_to_end_ms": 8837.11385,
    "end_to_end_ms_q25": 8768.282552,
    "end_to_end_ms_q75": 8852.740913499998,
    "output_tokens_per_s_e2e": 1.8105458718289569,
    "output_tokens_per_s_e2e_q25": 1.8073554860237997,
    "output_tokens_per_s_e2e_q75": 1.824871168771017,
    "all_tokens_per_s_e2e": 463.49974318821296,
    "all_tokens_per_s_e2e_q25": 462.6830044220927,
    "all_tokens_per_s_e2e_q75": 467.16701920538037,
    "energy_net_psys_j": 291.17361407593637,
    "energy_net_psys_j_q25": 290.45306624886246,
    "energy_net_psys_j_q75": 291.40285178421004,
    "prefill_energy_net_psys_j": 231.97102930961674,
    "prefill_energy_net_psys_j_q25": 228.48211675690027,
    "prefill_energy_net_psys_j_q75": 232.25588411777665,
    "decode_energy_net_psys_j": 59.09135056654719,
    "decode_energy_net_psys_j_q25": 58.426419839359525,
    "decode_energy_net_psys_j_q75": 62.635880219149875,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.9487988661592268,
    "energy_reduction_vs_gpu": 0.09146786679955576
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 4096,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 8327.051625,
    "ttft_ms_q25": 8321.200045,
    "ttft_ms_q75": 8330.5022765,
    "prefill_tokens_per_s": 489.9693413393484,
    "prefill_tokens_per_s_q25": 489.7664708099305,
    "prefill_tokens_per_s_q75": 490.31413684737333,
    "handoff_plus_first_decode_ms": 330.0977,
    "handoff_plus_first_decode_ms_q25": 316.156475,
    "handoff_plus_first_decode_ms_q75": 332.4441515,
    "steady_tpot_ms": 214.40785449999998,
    "steady_tpot_ms_q25": 214.3305785,
    "steady_tpot_ms_q75": 214.41199475,
    "steady_tpot_p95_ms": 215.12930295,
    "steady_tpot_p95_ms_q25": 215.090931,
    "steady_tpot_p95_ms_q75": 215.477234925,
    "decode_tokens_per_s": 4.503460418454397,
    "decode_tokens_per_s_q25": 4.49964235404687,
    "decode_tokens_per_s_q75": 4.518838452798773,
    "decode_ms": 3330.77203,
    "decode_ms_q25": 3319.475541,
    "decode_ms_q75": 3333.6006774999996,
    "end_to_end_ms": 11646.120495,
    "end_to_end_ms_q25": 11640.675586,
    "end_to_end_ms_q75": 11658.251374,
    "output_tokens_per_s_e2e": 1.373848055828483,
    "output_tokens_per_s_e2e_q25": 1.3724199976750873,
    "output_tokens_per_s_e2e_q75": 1.3744909719726897,
    "all_tokens_per_s_e2e": 351.7051022920916,
    "all_tokens_per_s_e2e_q25": 351.33951940482234,
    "all_tokens_per_s_e2e_q75": 351.86968882500855,
    "energy_net_psys_j": 297.0032745093148,
    "energy_net_psys_j_q25": 293.463646088685,
    "energy_net_psys_j_q75": 297.54450137030153,
    "prefill_energy_net_psys_j": 232.16252855588326,
    "prefill_energy_net_psys_j_q25": 231.49269659672274,
    "prefill_energy_net_psys_j_q75": 235.57845311031213,
    "decode_energy_net_psys_j": 59.09135056654719,
    "decode_energy_net_psys_j_q25": 58.426419839359525,
    "decode_energy_net_psys_j_q75": 62.635880219149875,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.7199516443780363,
    "energy_reduction_vs_gpu": 0.07327791560435493
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 4096,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 8327.051625,
    "ttft_ms_q25": 8321.200045,
    "ttft_ms_q75": 8330.5022765,
    "prefill_tokens_per_s": 489.9693413393484,
    "prefill_tokens_per_s_q25": 489.7664708099305,
    "prefill_tokens_per_s_q75": 490.31413684737333,
    "handoff_plus_first_decode_ms": 209.129247,
    "handoff_plus_first_decode_ms_q25": 207.3933045,
    "handoff_plus_first_decode_ms_q75": 210.9861555,
    "steady_tpot_ms": 188.47727250000003,
    "steady_tpot_ms_q25": 188.40073825000002,
    "steady_tpot_ms_q75": 188.802173,
    "steady_tpot_p95_ms": 190.72192264999998,
    "steady_tpot_p95_ms_q25": 190.219103625,
    "steady_tpot_p95_ms_q75": 191.32966362500002,
    "decode_tokens_per_s": 5.254028638369279,
    "decode_tokens_per_s_q25": 5.253332518057384,
    "decode_tokens_per_s_q75": 5.254524002663129,
    "decode_ms": 2854.9520820000007,
    "decode_ms_q25": 2854.68296,
    "decode_ms_q75": 2855.3304425,
    "end_to_end_ms": 11182.760428,
    "end_to_end_ms_q25": 11176.261365499999,
    "end_to_end_ms_q75": 11185.832719000002,
    "output_tokens_per_s_e2e": 1.4307737434791445,
    "output_tokens_per_s_e2e_q25": 1.430380876322689,
    "output_tokens_per_s_e2e_q75": 1.4316062310227453,
    "all_tokens_per_s_e2e": 366.278078330661,
    "all_tokens_per_s_e2e_q25": 366.1775043386084,
    "all_tokens_per_s_e2e_q75": 366.4911951418228,
    "energy_net_psys_j": 320.679463772334,
    "energy_net_psys_j_q25": 320.62188891325565,
    "energy_net_psys_j_q75": 324.5941031124986,
    "prefill_energy_net_psys_j": 232.16252855588326,
    "prefill_energy_net_psys_j_q25": 231.49269659672274,
    "prefill_energy_net_psys_j_q75": 235.57845311031213,
    "decode_energy_net_psys_j": 89.51436478792222,
    "decode_energy_net_psys_j_q25": 89.01565000218648,
    "decode_energy_net_psys_j_q75": 89.62790710226865,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.7497829945463266,
    "energy_reduction_vs_gpu": -0.0005975239867421767
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 8192,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 13534.217505999999,
    "ttft_ms_q25": 13441.503799999999,
    "ttft_ms_q75": 14256.6174665,
    "prefill_tokens_per_s": 604.0984634963498,
    "prefill_tokens_per_s_q25": 574.9643291451038,
    "prefill_tokens_per_s_q75": 608.2942150922063,
    "handoff_plus_first_decode_ms": 235.474582,
    "handoff_plus_first_decode_ms_q25": 235.0273355,
    "handoff_plus_first_decode_ms_q75": 249.6081695,
    "steady_tpot_ms": 196.93333750000002,
    "steady_tpot_ms_q25": 196.81365750000003,
    "steady_tpot_ms_q75": 197.04648200000003,
    "steady_tpot_p95_ms": 200.4805697,
    "steady_tpot_p95_ms_q25": 199.683263125,
    "steady_tpot_p95_ms_q75": 200.726778025,
    "decode_tokens_per_s": 4.994483062377571,
    "decode_tokens_per_s_q25": 4.9737249486687,
    "decode_tokens_per_s_q75": 5.002145834896865,
    "decode_ms": 3003.313819000001,
    "decode_ms_q25": 2998.7200885,
    "decode_ms_q75": 3015.9008465000015,
    "end_to_end_ms": 16528.343864,
    "end_to_end_ms_q25": 16452.810916000002,
    "end_to_end_ms_q75": 17255.337555,
    "output_tokens_per_s_e2e": 0.9680340711478799,
    "output_tokens_per_s_e2e_q25": 0.9288981742071483,
    "output_tokens_per_s_e2e_q75": 0.9724986999647753,
    "all_tokens_per_s_e2e": 495.6334444277145,
    "all_tokens_per_s_e2e_q25": 475.5958651940599,
    "all_tokens_per_s_e2e_q75": 497.919334381965,
    "energy_net_psys_j": 616.5810162520157,
    "energy_net_psys_j_q25": 611.4555356138947,
    "energy_net_psys_j_q75": 635.3122916265785,
    "prefill_energy_net_psys_j": 526.5659219765113,
    "prefill_energy_net_psys_j_q25": 520.2107352229555,
    "prefill_energy_net_psys_j_q75": 544.9391911780941,
    "decode_energy_net_psys_j": 90.73110662146432,
    "decode_energy_net_psys_j_q25": 90.37310044848437,
    "decode_energy_net_psys_j_q75": 91.60280656391917,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 8192,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 13534.217505999999,
    "ttft_ms_q25": 13441.503799999999,
    "ttft_ms_q75": 14256.6174665,
    "prefill_tokens_per_s": 604.0984634963498,
    "prefill_tokens_per_s_q25": 574.9643291451038,
    "prefill_tokens_per_s_q75": 608.2942150922063,
    "handoff_plus_first_decode_ms": 402.618993,
    "handoff_plus_first_decode_ms_q25": 402.290937,
    "handoff_plus_first_decode_ms_q75": 415.835902,
    "steady_tpot_ms": 224.522828,
    "steady_tpot_ms_q25": 224.4906115,
    "steady_tpot_ms_q75": 224.76623325,
    "steady_tpot_p95_ms": 225.88628919999996,
    "steady_tpot_p95_ms_q25": 225.71495657499997,
    "steady_tpot_p95_ms_q75": 226.15309994999998,
    "decode_tokens_per_s": 4.22430750449732,
    "decode_tokens_per_s_q25": 4.21039922053839,
    "decode_tokens_per_s_q75": 4.226318357939814,
    "decode_ms": 3550.8778619999975,
    "decode_ms_q25": 3549.189182,
    "decode_ms_q75": 3562.6464129999986,
    "end_to_end_ms": 17085.095367999995,
    "end_to_end_ms_q25": 17004.150212999997,
    "end_to_end_ms_q75": 17805.806648499998,
    "output_tokens_per_s_e2e": 0.9364887731307399,
    "output_tokens_per_s_e2e_q25": 0.9000578568925386,
    "output_tokens_per_s_e2e_q75": 0.9409680800136279,
    "all_tokens_per_s_e2e": 479.4822518429388,
    "all_tokens_per_s_e2e_q25": 460.82962272897976,
    "all_tokens_per_s_e2e_q75": 481.77565696697746,
    "energy_net_psys_j": 587.2137486507721,
    "energy_net_psys_j_q25": 583.1809150293245,
    "energy_net_psys_j_q75": 606.9761207902717,
    "prefill_energy_net_psys_j": 526.5659219765113,
    "prefill_energy_net_psys_j_q25": 520.2107352229555,
    "prefill_energy_net_psys_j_q75": 544.9391911780941,
    "decode_energy_net_psys_j": 63.42603255009427,
    "decode_energy_net_psys_j_q25": 62.03692961217753,
    "decode_energy_net_psys_j_q75": 64.35928274428575,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.9674130291925218,
    "energy_reduction_vs_gpu": 0.047629211453439124
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 8192,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 19051.627731,
    "ttft_ms_q25": 19030.7399085,
    "ttft_ms_q75": 19067.34142,
    "prefill_tokens_per_s": 429.14968292690077,
    "prefill_tokens_per_s_q25": 428.7963053040074,
    "prefill_tokens_per_s_q75": 429.6212280163815,
    "handoff_plus_first_decode_ms": 402.618993,
    "handoff_plus_first_decode_ms_q25": 402.290937,
    "handoff_plus_first_decode_ms_q75": 415.835902,
    "steady_tpot_ms": 224.522828,
    "steady_tpot_ms_q25": 224.4906115,
    "steady_tpot_ms_q75": 224.76623325,
    "steady_tpot_p95_ms": 225.88628919999996,
    "steady_tpot_p95_ms_q25": 225.71495657499997,
    "steady_tpot_p95_ms_q75": 226.15309994999998,
    "decode_tokens_per_s": 4.22430750449732,
    "decode_tokens_per_s_q25": 4.21039922053839,
    "decode_tokens_per_s_q75": 4.226318357939814,
    "decode_ms": 3550.8778619999975,
    "decode_ms_q25": 3549.189182,
    "decode_ms_q75": 3562.6464129999986,
    "end_to_end_ms": 22599.128233000003,
    "end_to_end_ms_q25": 22591.697641500003,
    "end_to_end_ms_q75": 22616.530602,
    "output_tokens_per_s_e2e": 0.7079919116807464,
    "output_tokens_per_s_e2e_q25": 0.7074475636652305,
    "output_tokens_per_s_e2e_q75": 0.708224852544816,
    "all_tokens_per_s_e2e": 362.49185878054215,
    "all_tokens_per_s_e2e_q25": 362.21315259659804,
    "all_tokens_per_s_e2e_q75": 362.6111245029458,
    "energy_net_psys_j": 593.999688863384,
    "energy_net_psys_j_q25": 589.3421377157467,
    "energy_net_psys_j_q75": 598.0619848365291,
    "prefill_energy_net_psys_j": 528.7071559249067,
    "prefill_energy_net_psys_j_q25": 526.3719579093777,
    "prefill_energy_net_psys_j_q75": 533.7027020922433,
    "decode_energy_net_psys_j": 63.42603255009427,
    "decode_energy_net_psys_j_q25": 62.03692961217753,
    "decode_energy_net_psys_j_q75": 64.35928274428575,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.7313708605743808,
    "energy_reduction_vs_gpu": 0.036623455463964616
  },
  {
    "precision": "FP16",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 8192,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 19051.627731,
    "ttft_ms_q25": 19030.7399085,
    "ttft_ms_q75": 19067.34142,
    "prefill_tokens_per_s": 429.14968292690077,
    "prefill_tokens_per_s_q25": 428.7963053040074,
    "prefill_tokens_per_s_q75": 429.6212280163815,
    "handoff_plus_first_decode_ms": 235.474582,
    "handoff_plus_first_decode_ms_q25": 235.0273355,
    "handoff_plus_first_decode_ms_q75": 249.6081695,
    "steady_tpot_ms": 196.93333750000002,
    "steady_tpot_ms_q25": 196.81365750000003,
    "steady_tpot_ms_q75": 197.04648200000003,
    "steady_tpot_p95_ms": 200.4805697,
    "steady_tpot_p95_ms_q25": 199.683263125,
    "steady_tpot_p95_ms_q75": 200.726778025,
    "decode_tokens_per_s": 4.994483062377571,
    "decode_tokens_per_s_q25": 4.9737249486687,
    "decode_tokens_per_s_q75": 5.002145834896865,
    "decode_ms": 3003.313819000001,
    "decode_ms_q25": 2998.7200885,
    "decode_ms_q75": 3015.9008465000015,
    "end_to_end_ms": 22054.941550000003,
    "end_to_end_ms_q25": 22046.640755,
    "end_to_end_ms_q75": 22066.061508500003,
    "output_tokens_per_s_e2e": 0.7254610021852449,
    "output_tokens_per_s_e2e_q25": 0.7250955979215834,
    "output_tokens_per_s_e2e_q75": 0.7257342488576606,
    "all_tokens_per_s_e2e": 371.4360331188454,
    "all_tokens_per_s_e2e_q25": 371.2489461358507,
    "all_tokens_per_s_e2e_q75": 371.5759354151222,
    "energy_net_psys_j": 621.1816624312808,
    "energy_net_psys_j_q25": 617.6167583003169,
    "energy_net_psys_j_q75": 625.3055086561624,
    "prefill_energy_net_psys_j": 528.7071559249067,
    "prefill_energy_net_psys_j_q25": 526.3719579093777,
    "prefill_energy_net_psys_j_q75": 533.7027020922433,
    "decode_energy_net_psys_j": 90.73110662146432,
    "decode_energy_net_psys_j_q25": 90.37310044848437,
    "decode_energy_net_psys_j_q75": 91.60280656391917,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.7494168065024863,
    "energy_reduction_vs_gpu": -0.007461543670661097
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 1024,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 258.676058,
    "ttft_ms_q25": 249.456735,
    "ttft_ms_q75": 261.883168,
    "prefill_tokens_per_s": 3896.765737786216,
    "prefill_tokens_per_s_q25": 3849.621962995231,
    "prefill_tokens_per_s_q75": 4046.3075668165106,
    "handoff_plus_first_decode_ms": 32.335183,
    "handoff_plus_first_decode_ms_q25": 28.476062,
    "handoff_plus_first_decode_ms_q75": 36.0558015,
    "steady_tpot_ms": 18.722242,
    "steady_tpot_ms_q25": 18.680516750000002,
    "steady_tpot_ms_q75": 18.731333499999998,
    "steady_tpot_p95_ms": 20.45485955,
    "steady_tpot_p95_ms_q25": 20.2223636,
    "steady_tpot_p95_ms_q75": 23.408128325,
    "decode_tokens_per_s": 49.30111868083774,
    "decode_tokens_per_s_q25": 48.35031790623841,
    "decode_tokens_per_s_q75": 50.503898346549306,
    "decode_ms": 304.252731,
    "decode_ms_q25": 297.17532800000004,
    "decode_ms_q75": 310.35582550000004,
    "end_to_end_ms": 548.773983,
    "end_to_end_ms_q25": 546.632063,
    "end_to_end_ms_q75": 565.1615905000001,
    "output_tokens_per_s_e2e": 29.15590114628302,
    "output_tokens_per_s_e2e_q25": 28.33431035751181,
    "output_tokens_per_s_e2e_q75": 29.27059488057023,
    "all_tokens_per_s_e2e": 1865.9776733621134,
    "all_tokens_per_s_e2e_q25": 1813.3958628807559,
    "all_tokens_per_s_e2e_q75": 1873.3180723564947,
    "energy_net_psys_j": 19.008366268236916,
    "energy_net_psys_j_q25": 18.873820535060823,
    "energy_net_psys_j_q75": 19.184268749006918,
    "prefill_energy_net_psys_j": 7.7984988544351985,
    "prefill_energy_net_psys_j_q25": 7.4164487334728095,
    "prefill_energy_net_psys_j_q75": 7.816856745044378,
    "decode_energy_net_psys_j": 11.52495659412336,
    "decode_energy_net_psys_j_q25": 11.232866270786444,
    "decode_energy_net_psys_j_q75": 11.749462124924928,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 1024,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 258.676058,
    "ttft_ms_q25": 249.456735,
    "ttft_ms_q75": 261.883168,
    "prefill_tokens_per_s": 3896.765737786216,
    "prefill_tokens_per_s_q25": 3849.621962995231,
    "prefill_tokens_per_s_q75": 4046.3075668165106,
    "handoff_plus_first_decode_ms": 428.263869,
    "handoff_plus_first_decode_ms_q25": 427.56293949999997,
    "handoff_plus_first_decode_ms_q75": 512.7273015000001,
    "steady_tpot_ms": 422.7833595,
    "steady_tpot_ms_q25": 422.41823225,
    "steady_tpot_ms_q75": 422.849783,
    "steady_tpot_p95_ms": 423.48913335,
    "steady_tpot_p95_ms_q25": 423.4693372,
    "steady_tpot_p95_ms_q75": 424.7582149,
    "decode_tokens_per_s": 2.3637284280508135,
    "decode_tokens_per_s_q25": 2.331620801672016,
    "decode_tokens_per_s_q75": 2.365490507573279,
    "decode_ms": 6345.906671,
    "decode_ms_q25": 6341.1830549999995,
    "decode_ms_q75": 6434.513251999999,
    "end_to_end_ms": 6610.996948999999,
    "end_to_end_ms_q25": 6603.066223,
    "end_to_end_ms_q75": 6687.177097,
    "output_tokens_per_s_e2e": 2.4202098599395376,
    "output_tokens_per_s_e2e_q25": 2.392949441978095,
    "output_tokens_per_s_e2e_q75": 2.4231201900260704,
    "all_tokens_per_s_e2e": 154.8934310361304,
    "all_tokens_per_s_e2e_q25": 153.14876428659807,
    "all_tokens_per_s_e2e_q75": 155.0796921616685,
    "energy_net_psys_j": 94.22866893801896,
    "energy_net_psys_j_q25": 93.15838990974521,
    "energy_net_psys_j_q75": 96.40623733088943,
    "prefill_energy_net_psys_j": 7.7984988544351985,
    "prefill_energy_net_psys_j_q25": 7.4164487334728095,
    "prefill_energy_net_psys_j_q75": 7.816856745044378,
    "decode_energy_net_psys_j": 86.3934543023654,
    "decode_energy_net_psys_j_q25": 85.34153316470083,
    "decode_energy_net_psys_j_q75": 88.97143070680744,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.08300926278343078,
    "energy_reduction_vs_gpu": -3.9572208157349946
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 1024,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 715.7900470000001,
    "ttft_ms_q25": 715.349371,
    "ttft_ms_q75": 716.878866,
    "prefill_tokens_per_s": 1408.2341661842079,
    "prefill_tokens_per_s_q25": 1406.0985378702533,
    "prefill_tokens_per_s_q75": 1409.1022140954415,
    "handoff_plus_first_decode_ms": 428.263869,
    "handoff_plus_first_decode_ms_q25": 427.56293949999997,
    "handoff_plus_first_decode_ms_q75": 512.7273015000001,
    "steady_tpot_ms": 422.7833595,
    "steady_tpot_ms_q25": 422.41823225,
    "steady_tpot_ms_q75": 422.849783,
    "steady_tpot_p95_ms": 423.48913335,
    "steady_tpot_p95_ms_q25": 423.4693372,
    "steady_tpot_p95_ms_q75": 424.7582149,
    "decode_tokens_per_s": 2.3637284280508135,
    "decode_tokens_per_s_q25": 2.331620801672016,
    "decode_tokens_per_s_q75": 2.365490507573279,
    "decode_ms": 6345.906671,
    "decode_ms_q25": 6341.1830549999995,
    "decode_ms_q75": 6434.513251999999,
    "end_to_end_ms": 7060.815366,
    "end_to_end_ms_q25": 7057.621244999999,
    "end_to_end_ms_q75": 7149.862622999999,
    "output_tokens_per_s_e2e": 2.2660272462363094,
    "output_tokens_per_s_e2e_q25": 2.2381523985161986,
    "output_tokens_per_s_e2e_q75": 2.267053263669979,
    "all_tokens_per_s_e2e": 145.0257437591238,
    "all_tokens_per_s_e2e_q25": 143.2417535050367,
    "all_tokens_per_s_e2e_q75": 145.09140887487865,
    "energy_net_psys_j": 97.75641667800228,
    "energy_net_psys_j_q25": 96.67973852350553,
    "energy_net_psys_j_q75": 100.58658423217739,
    "prefill_energy_net_psys_j": 11.362962375636888,
    "prefill_energy_net_psys_j_q25": 11.338205358804714,
    "prefill_energy_net_psys_j_q75": 11.615153525369948,
    "decode_energy_net_psys_j": 86.3934543023654,
    "decode_energy_net_psys_j_q25": 85.34153316470083,
    "decode_energy_net_psys_j_q75": 88.97143070680744,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.07772104984397635,
    "energy_reduction_vs_gpu": -4.1428100289372995
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 1024,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 715.7900470000001,
    "ttft_ms_q25": 715.349371,
    "ttft_ms_q75": 716.878866,
    "prefill_tokens_per_s": 1408.2341661842079,
    "prefill_tokens_per_s_q25": 1406.0985378702533,
    "prefill_tokens_per_s_q75": 1409.1022140954415,
    "handoff_plus_first_decode_ms": 32.335183,
    "handoff_plus_first_decode_ms_q25": 28.476062,
    "handoff_plus_first_decode_ms_q75": 36.0558015,
    "steady_tpot_ms": 18.722242,
    "steady_tpot_ms_q25": 18.680516750000002,
    "steady_tpot_ms_q75": 18.731333499999998,
    "steady_tpot_p95_ms": 20.45485955,
    "steady_tpot_p95_ms_q25": 20.2223636,
    "steady_tpot_p95_ms_q75": 23.408128325,
    "decode_tokens_per_s": 49.30111868083774,
    "decode_tokens_per_s_q25": 48.35031790623841,
    "decode_tokens_per_s_q75": 50.503898346549306,
    "decode_ms": 304.252731,
    "decode_ms_q25": 297.17532800000004,
    "decode_ms_q75": 310.35582550000004,
    "end_to_end_ms": 1020.042778,
    "end_to_end_ms_q25": 1014.054194,
    "end_to_end_ms_q75": 1025.7051965,
    "output_tokens_per_s_e2e": 15.685616667343336,
    "output_tokens_per_s_e2e_q25": 15.599499433575943,
    "output_tokens_per_s_e2e_q75": 15.778799722168692,
    "all_tokens_per_s_e2e": 1003.8794667099735,
    "all_tokens_per_s_e2e_q25": 998.3679637488603,
    "all_tokens_per_s_e2e_q75": 1009.8431822187963,
    "energy_net_psys_j": 22.88791896976025,
    "energy_net_psys_j_q25": 22.57107162959116,
    "energy_net_psys_j_q75": 23.364615650294876,
    "prefill_energy_net_psys_j": 11.362962375636888,
    "prefill_energy_net_psys_j_q25": 11.338205358804714,
    "prefill_energy_net_psys_j_q75": 11.615153525369948,
    "decode_energy_net_psys_j": 11.52495659412336,
    "decode_energy_net_psys_j_q25": 11.232866270786444,
    "decode_energy_net_psys_j_q75": 11.749462124924928,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.5379911458968244,
    "energy_reduction_vs_gpu": -0.20409711422733312
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 2048,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 563.861085,
    "ttft_ms_q25": 540.7457489999999,
    "ttft_ms_q75": 563.9430460000001,
    "prefill_tokens_per_s": 3603.724488275335,
    "prefill_tokens_per_s_q25": 3603.200814929086,
    "prefill_tokens_per_s_q75": 3764.652633735583,
    "handoff_plus_first_decode_ms": 30.489342,
    "handoff_plus_first_decode_ms_q25": 30.156737,
    "handoff_plus_first_decode_ms_q75": 37.794177000000005,
    "steady_tpot_ms": 19.3590285,
    "steady_tpot_ms_q25": 19.310399500000003,
    "steady_tpot_ms_q75": 19.42046825,
    "steady_tpot_p95_ms": 21.360242849999995,
    "steady_tpot_p95_ms_q25": 20.723510174999998,
    "steady_tpot_p95_ms_q75": 21.909633949999996,
    "decode_tokens_per_s": 48.71870171753366,
    "decode_tokens_per_s_q25": 47.52197164148824,
    "decode_tokens_per_s_q75": 49.14911932672539,
    "decode_ms": 307.8899779999999,
    "decode_ms_q25": 305.21707549999996,
    "decode_ms_q75": 315.84376699999996,
    "end_to_end_ms": 866.5691800000001,
    "end_to_end_ms_q25": 853.9985745,
    "end_to_end_ms_q75": 869.1601215000001,
    "output_tokens_per_s_e2e": 18.463615334207937,
    "output_tokens_per_s_e2e_q25": 18.408739407765353,
    "output_tokens_per_s_e2e_q75": 18.739454542103438,
    "all_tokens_per_s_e2e": 2363.342762778616,
    "all_tokens_per_s_e2e_q25": 2356.318644193965,
    "all_tokens_per_s_e2e_q75": 2398.65018138924,
    "energy_net_psys_j": 28.135222944952325,
    "energy_net_psys_j_q25": 28.004517460824047,
    "energy_net_psys_j_q75": 28.28943526997993,
    "prefill_energy_net_psys_j": 16.986121113617084,
    "prefill_energy_net_psys_j_q25": 16.25691982914457,
    "prefill_energy_net_psys_j_q75": 17.177019131899208,
    "decode_energy_net_psys_j": 11.14910183133524,
    "decode_energy_net_psys_j_q25": 11.112416138080722,
    "decode_energy_net_psys_j_q75": 11.747597631679474,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 2048,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 563.861085,
    "ttft_ms_q25": 540.7457489999999,
    "ttft_ms_q75": 563.9430460000001,
    "prefill_tokens_per_s": 3603.724488275335,
    "prefill_tokens_per_s_q25": 3603.200814929086,
    "prefill_tokens_per_s_q75": 3764.652633735583,
    "handoff_plus_first_decode_ms": 455.564631,
    "handoff_plus_first_decode_ms_q25": 454.43201650000003,
    "handoff_plus_first_decode_ms_q75": 531.5957195,
    "steady_tpot_ms": 424.542054,
    "steady_tpot_ms_q25": 424.36665475,
    "steady_tpot_ms_q75": 424.65971775,
    "steady_tpot_p95_ms": 425.34586365,
    "steady_tpot_p95_ms_q25": 425.267551475,
    "steady_tpot_p95_ms_q75": 426.050509475,
    "decode_tokens_per_s": 2.343448505097232,
    "decode_tokens_per_s_q25": 2.3162625481922383,
    "decode_tokens_per_s_q75": 2.3454854728912924,
    "decode_ms": 6400.823388000001,
    "decode_ms_q25": 6395.269332,
    "decode_ms_q75": 6476.8420365,
    "end_to_end_ms": 6964.684473000001,
    "end_to_end_ms_q25": 6959.212378,
    "end_to_end_ms_q75": 7017.587785500001,
    "output_tokens_per_s_e2e": 2.2973043591604494,
    "output_tokens_per_s_e2e_q25": 2.280115311223644,
    "output_tokens_per_s_e2e_q75": 2.29911217296793,
    "all_tokens_per_s_e2e": 294.0549579725375,
    "all_tokens_per_s_e2e_q25": 291.85475983662644,
    "all_tokens_per_s_e2e_q75": 294.286358139895,
    "energy_net_psys_j": 101.01422013008192,
    "energy_net_psys_j_q25": 99.63349111381083,
    "energy_net_psys_j_q75": 102.53193960521517,
    "prefill_energy_net_psys_j": 16.986121113617084,
    "prefill_energy_net_psys_j_q25": 16.25691982914457,
    "prefill_energy_net_psys_j_q75": 17.177019131899208,
    "decode_energy_net_psys_j": 83.6463029799006,
    "decode_energy_net_psys_j_q25": 82.45647198191163,
    "decode_energy_net_psys_j_q75": 86.0841217577885,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.12442332217050601,
    "energy_reduction_vs_gpu": -2.5903117003096168
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 2048,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1425.810721,
    "ttft_ms_q25": 1424.4794295000002,
    "ttft_ms_q75": 1427.280946,
    "prefill_tokens_per_s": 1425.1541036069964,
    "prefill_tokens_per_s_q25": 1423.6875800901623,
    "prefill_tokens_per_s_q75": 1426.4872715981996,
    "handoff_plus_first_decode_ms": 455.564631,
    "handoff_plus_first_decode_ms_q25": 454.43201650000003,
    "handoff_plus_first_decode_ms_q75": 531.5957195,
    "steady_tpot_ms": 424.542054,
    "steady_tpot_ms_q25": 424.36665475,
    "steady_tpot_ms_q75": 424.65971775,
    "steady_tpot_p95_ms": 425.34586365,
    "steady_tpot_p95_ms_q25": 425.267551475,
    "steady_tpot_p95_ms_q75": 426.050509475,
    "decode_tokens_per_s": 2.343448505097232,
    "decode_tokens_per_s_q25": 2.3162625481922383,
    "decode_tokens_per_s_q75": 2.3454854728912924,
    "decode_ms": 6400.823388000001,
    "decode_ms_q25": 6395.269332,
    "decode_ms_q75": 6476.8420365,
    "end_to_end_ms": 7823.971526000001,
    "end_to_end_ms_q25": 7821.2189865,
    "end_to_end_ms_q75": 7901.321466000001,
    "output_tokens_per_s_e2e": 2.0449972174400264,
    "output_tokens_per_s_e2e_q25": 2.025171809922905,
    "output_tokens_per_s_e2e_q75": 2.0457171713666122,
    "all_tokens_per_s_e2e": 261.7596438323234,
    "all_tokens_per_s_e2e_q25": 259.22199167013184,
    "all_tokens_per_s_e2e_q75": 261.85179793492637,
    "energy_net_psys_j": 107.6461381745534,
    "energy_net_psys_j_q25": 105.79259181837244,
    "energy_net_psys_j_q75": 110.44493746447417,
    "prefill_energy_net_psys_j": 23.999835194652803,
    "prefill_energy_net_psys_j_q25": 23.33611983646081,
    "prefill_energy_net_psys_j_q75": 24.360815706685678,
    "decode_energy_net_psys_j": 83.6463029799006,
    "decode_energy_net_psys_j_q25": 82.45647198191163,
    "decode_energy_net_psys_j_q75": 86.0841217577885,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.11075822261370535,
    "energy_reduction_vs_gpu": -2.8260275521955993
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 2048,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1425.810721,
    "ttft_ms_q25": 1424.4794295000002,
    "ttft_ms_q75": 1427.280946,
    "prefill_tokens_per_s": 1425.1541036069964,
    "prefill_tokens_per_s_q25": 1423.6875800901623,
    "prefill_tokens_per_s_q75": 1426.4872715981996,
    "handoff_plus_first_decode_ms": 30.489342,
    "handoff_plus_first_decode_ms_q25": 30.156737,
    "handoff_plus_first_decode_ms_q75": 37.794177000000005,
    "steady_tpot_ms": 19.3590285,
    "steady_tpot_ms_q25": 19.310399500000003,
    "steady_tpot_ms_q75": 19.42046825,
    "steady_tpot_p95_ms": 21.360242849999995,
    "steady_tpot_p95_ms_q25": 20.723510174999998,
    "steady_tpot_p95_ms_q75": 21.909633949999996,
    "decode_tokens_per_s": 48.71870171753366,
    "decode_tokens_per_s_q25": 47.52197164148824,
    "decode_tokens_per_s_q75": 49.14911932672539,
    "decode_ms": 307.8899779999999,
    "decode_ms_q25": 305.21707549999996,
    "decode_ms_q75": 315.84376699999996,
    "end_to_end_ms": 1731.2953440000001,
    "end_to_end_ms_q25": 1731.16673,
    "end_to_end_ms_q75": 1740.4518105000002,
    "output_tokens_per_s_e2e": 9.241635204212736,
    "output_tokens_per_s_e2e_q25": 9.193269679046331,
    "output_tokens_per_s_e2e_q75": 9.242321846326025,
    "all_tokens_per_s_e2e": 1182.9293061392302,
    "all_tokens_per_s_e2e_q25": 1176.7385189179304,
    "all_tokens_per_s_e2e_q75": 1183.0171963297312,
    "energy_net_psys_j": 35.07556563947901,
    "energy_net_psys_j_q25": 34.448535974541535,
    "energy_net_psys_j_q75": 36.071727645110634,
    "prefill_energy_net_psys_j": 23.999835194652803,
    "prefill_energy_net_psys_j_q25": 23.33611983646081,
    "prefill_energy_net_psys_j_q75": 24.360815706685678,
    "decode_energy_net_psys_j": 11.14910183133524,
    "decode_energy_net_psys_j_q25": 11.112416138080722,
    "decode_energy_net_psys_j_q75": 11.747597631679474,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.5005322650483602,
    "energy_reduction_vs_gpu": -0.24667807708884126
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 4096,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1428.511305,
    "ttft_ms_q25": 1382.6912465,
    "ttft_ms_q75": 1429.152221,
    "prefill_tokens_per_s": 2856.12020410297,
    "prefill_tokens_per_s_q25": 2854.839925796016,
    "prefill_tokens_per_s_q75": 2954.011157255167,
    "handoff_plus_first_decode_ms": 46.114875,
    "handoff_plus_first_decode_ms_q25": 40.593903499999996,
    "handoff_plus_first_decode_ms_q75": 54.083431000000004,
    "steady_tpot_ms": 21.2911095,
    "steady_tpot_ms_q25": 21.28367175,
    "steady_tpot_ms_q75": 21.34767375,
    "steady_tpot_p95_ms": 30.2459046,
    "steady_tpot_p95_ms_q25": 27.58292205,
    "steady_tpot_p95_ms_q75": 30.443447999999997,
    "decode_tokens_per_s": 40.425453535386616,
    "decode_tokens_per_s_q25": 39.854156287275416,
    "decode_tokens_per_s_q75": 42.04361476601651,
    "decode_ms": 371.053351,
    "decode_ms_q25": 357.3016395000001,
    "decode_ms_q75": 376.4496425000001,
    "end_to_end_ms": 1773.3430650000003,
    "end_to_end_ms_q25": 1746.0300935,
    "end_to_end_ms_q75": 1786.4538605000002,
    "output_tokens_per_s_e2e": 9.022506877427013,
    "output_tokens_per_s_e2e_q25": 8.956773066656233,
    "output_tokens_per_s_e2e_q75": 9.165887931846811,
    "all_tokens_per_s_e2e": 2309.7617606213153,
    "all_tokens_per_s_e2e_q25": 2292.9339050639956,
    "all_tokens_per_s_e2e_q75": 2346.4673105527836,
    "energy_net_psys_j": 54.844003295893515,
    "energy_net_psys_j_q25": 53.97734479587888,
    "energy_net_psys_j_q75": 54.970716725069366,
    "prefill_energy_net_psys_j": 42.46399896559708,
    "prefill_energy_net_psys_j_q25": 40.855619776760605,
    "prefill_energy_net_psys_j_q75": 42.49913510894337,
    "decode_energy_net_psys_j": 12.633431188648146,
    "decode_energy_net_psys_j_q25": 12.471581616125999,
    "decode_energy_net_psys_j_q75": 13.24843844829413,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 4096,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1428.511305,
    "ttft_ms_q25": 1382.6912465,
    "ttft_ms_q75": 1429.152221,
    "prefill_tokens_per_s": 2856.12020410297,
    "prefill_tokens_per_s_q25": 2854.839925796016,
    "prefill_tokens_per_s_q75": 2954.011157255167,
    "handoff_plus_first_decode_ms": 503.12422599999996,
    "handoff_plus_first_decode_ms_q25": 499.85916699999996,
    "handoff_plus_first_decode_ms_q75": 589.86765,
    "steady_tpot_ms": 428.0453165,
    "steady_tpot_ms_q25": 427.7648865,
    "steady_tpot_ms_q75": 428.14940275000004,
    "steady_tpot_p95_ms": 429.79907105,
    "steady_tpot_p95_ms_q25": 428.888505925,
    "steady_tpot_p95_ms_q75": 430.2993381,
    "decode_tokens_per_s": 2.3094296365368603,
    "decode_tokens_per_s_q25": 2.2784511735904087,
    "decode_tokens_per_s_q75": 2.31190794654563,
    "decode_ms": 6495.110204999999,
    "decode_ms_q25": 6488.155058,
    "decode_ms_q75": 6584.636778499999,
    "end_to_end_ms": 7923.621509999999,
    "end_to_end_ms_q25": 7917.307279000001,
    "end_to_end_ms_q75": 7967.328024999999,
    "output_tokens_per_s_e2e": 2.0192786820782915,
    "output_tokens_per_s_e2e_q25": 2.0082619234353514,
    "output_tokens_per_s_e2e_q75": 2.020890387710524,
    "all_tokens_per_s_e2e": 516.9353426120426,
    "all_tokens_per_s_e2e_q25": 514.11505239945,
    "all_tokens_per_s_e2e_q75": 517.3479392538942,
    "energy_net_psys_j": 126.08822509118554,
    "energy_net_psys_j_q25": 125.42869198447498,
    "energy_net_psys_j_q75": 126.11592997526176,
    "prefill_energy_net_psys_j": 42.46399896559708,
    "prefill_energy_net_psys_j_q25": 40.855619776760605,
    "prefill_energy_net_psys_j_q75": 42.49913510894337,
    "decode_energy_net_psys_j": 83.62422612558846,
    "decode_energy_net_psys_j_q25": 82.9295568755316,
    "decode_energy_net_psys_j_q75": 85.26031019850117,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.2238046154478674,
    "energy_reduction_vs_gpu": -1.2990339419775085
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 4096,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3119.049644,
    "ttft_ms_q25": 3117.103102,
    "ttft_ms_q75": 3120.2194215,
    "prefill_tokens_per_s": 1308.0907538129586,
    "prefill_tokens_per_s_q25": 1307.6005313401383,
    "prefill_tokens_per_s_q75": 1308.9081296151264,
    "handoff_plus_first_decode_ms": 503.12422599999996,
    "handoff_plus_first_decode_ms_q25": 499.85916699999996,
    "handoff_plus_first_decode_ms_q75": 589.86765,
    "steady_tpot_ms": 428.0453165,
    "steady_tpot_ms_q25": 427.7648865,
    "steady_tpot_ms_q75": 428.14940275000004,
    "steady_tpot_p95_ms": 429.79907105,
    "steady_tpot_p95_ms_q25": 428.888505925,
    "steady_tpot_p95_ms_q75": 430.2993381,
    "decode_tokens_per_s": 2.3094296365368603,
    "decode_tokens_per_s_q25": 2.2784511735904087,
    "decode_tokens_per_s_q75": 2.31190794654563,
    "decode_ms": 6495.110204999999,
    "decode_ms_q25": 6488.155058,
    "decode_ms_q75": 6584.636778499999,
    "end_to_end_ms": 9610.266764999998,
    "end_to_end_ms_q25": 9606.4279375,
    "end_to_end_ms_q75": 9701.7398805,
    "output_tokens_per_s_e2e": 1.6648861463732743,
    "output_tokens_per_s_e2e_q25": 1.6493353437078206,
    "output_tokens_per_s_e2e_q75": 1.6655517179862893,
    "all_tokens_per_s_e2e": 426.2108534715582,
    "all_tokens_per_s_e2e_q25": 422.2298479892021,
    "all_tokens_per_s_e2e_q75": 426.38123980449006,
    "energy_net_psys_j": 137.66066928337548,
    "energy_net_psys_j_q25": 137.0624293163785,
    "energy_net_psys_j_q75": 138.7169977022757,
    "prefill_energy_net_psys_j": 54.03644315778702,
    "prefill_energy_net_psys_j_q25": 53.45668750377452,
    "prefill_energy_net_psys_j_q75": 54.13287244084691,
    "decode_energy_net_psys_j": 83.62422612558846,
    "decode_energy_net_psys_j_q25": 82.9295568755316,
    "decode_energy_net_psys_j_q75": 85.26031019850117,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.18452589385535131,
    "energy_reduction_vs_gpu": -1.5100404968738474
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 4096,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3119.049644,
    "ttft_ms_q25": 3117.103102,
    "ttft_ms_q75": 3120.2194215,
    "prefill_tokens_per_s": 1308.0907538129586,
    "prefill_tokens_per_s_q25": 1307.6005313401383,
    "prefill_tokens_per_s_q75": 1308.9081296151264,
    "handoff_plus_first_decode_ms": 46.114875,
    "handoff_plus_first_decode_ms_q25": 40.593903499999996,
    "handoff_plus_first_decode_ms_q75": 54.083431000000004,
    "steady_tpot_ms": 21.2911095,
    "steady_tpot_ms_q25": 21.28367175,
    "steady_tpot_ms_q75": 21.34767375,
    "steady_tpot_p95_ms": 30.2459046,
    "steady_tpot_p95_ms_q25": 27.58292205,
    "steady_tpot_p95_ms_q75": 30.443447999999997,
    "decode_tokens_per_s": 40.425453535386616,
    "decode_tokens_per_s_q25": 39.854156287275416,
    "decode_tokens_per_s_q75": 42.04361476601651,
    "decode_ms": 371.053351,
    "decode_ms_q25": 357.3016395000001,
    "decode_ms_q75": 376.4496425000001,
    "end_to_end_ms": 3486.209911,
    "end_to_end_ms_q25": 3475.574519,
    "end_to_end_ms_q75": 3493.5527445000002,
    "output_tokens_per_s_e2e": 4.589511362903127,
    "output_tokens_per_s_e2e_q25": 4.57988524951776,
    "output_tokens_per_s_e2e_q75": 4.603598551925462,
    "all_tokens_per_s_e2e": 1174.9149089032005,
    "all_tokens_per_s_e2e_q25": 1172.4506238765466,
    "all_tokens_per_s_e2e_q75": 1178.5212292929182,
    "energy_net_psys_j": 66.66987434643517,
    "energy_net_psys_j_q25": 66.60445405697291,
    "energy_net_psys_j_q75": 66.70512595206864,
    "prefill_energy_net_psys_j": 54.03644315778702,
    "prefill_energy_net_psys_j_q25": 53.45668750377452,
    "prefill_energy_net_psys_j_q75": 54.13287244084691,
    "decode_energy_net_psys_j": 12.633431188648146,
    "decode_energy_net_psys_j_q25": 12.471581616125999,
    "decode_energy_net_psys_j_q75": 13.24843844829413,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.5086736341964351,
    "energy_reduction_vs_gpu": -0.21562742214019082
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 8192,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3931.624254,
    "ttft_ms_q25": 3847.9403665,
    "ttft_ms_q75": 3935.3583015,
    "prefill_tokens_per_s": 2079.54765557309,
    "prefill_tokens_per_s_q25": 2077.5763563405662,
    "prefill_tokens_per_s_q75": 2125.778467163829,
    "handoff_plus_first_decode_ms": 52.534369,
    "handoff_plus_first_decode_ms_q25": 50.268923,
    "handoff_plus_first_decode_ms_q75": 56.2366885,
    "steady_tpot_ms": 25.031965,
    "steady_tpot_ms_q25": 25.030341999999997,
    "steady_tpot_ms_q75": 25.149005000000002,
    "steady_tpot_p95_ms": 27.1545193,
    "steady_tpot_p95_ms_q25": 26.919350299999998,
    "steady_tpot_p95_ms_q75": 28.1688526,
    "decode_tokens_per_s": 36.71694083202306,
    "decode_tokens_per_s_q25": 36.34434649602082,
    "decode_tokens_per_s_q75": 36.752512545544135,
    "decode_ms": 408.53076700000065,
    "decode_ms_q25": 408.13574400000016,
    "decode_ms_q75": 412.76231650000045,
    "end_to_end_ms": 4340.1550210000005,
    "end_to_end_ms_q25": 4260.702683,
    "end_to_end_ms_q75": 4343.4940455,
    "output_tokens_per_s_e2e": 3.686504265995894,
    "output_tokens_per_s_e2e_q25": 3.683672473777217,
    "output_tokens_per_s_e2e_q75": 3.7565554130960916,
    "all_tokens_per_s_e2e": 1887.4901841898977,
    "all_tokens_per_s_e2e_q25": 1886.0403065739351,
    "all_tokens_per_s_e2e_q75": 1923.356371505199,
    "energy_net_psys_j": 129.08153154364902,
    "energy_net_psys_j_q25": 125.66201705239604,
    "energy_net_psys_j_q75": 129.60250391913198,
    "prefill_energy_net_psys_j": 114.36186444242055,
    "prefill_energy_net_psys_j_q25": 110.84935359375453,
    "prefill_energy_net_psys_j_q75": 114.86505231196605,
    "decode_energy_net_psys_j": 14.755236113103393,
    "decode_energy_net_psys_j_q25": 14.73745160716593,
    "decode_energy_net_psys_j_q75": 14.830447964578973,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 8192,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3931.624254,
    "ttft_ms_q25": 3847.9403665,
    "ttft_ms_q75": 3935.3583015,
    "prefill_tokens_per_s": 2079.54765557309,
    "prefill_tokens_per_s_q25": 2077.5763563405662,
    "prefill_tokens_per_s_q75": 2125.778467163829,
    "handoff_plus_first_decode_ms": 569.7060819999999,
    "handoff_plus_first_decode_ms_q25": 551.0542674999999,
    "handoff_plus_first_decode_ms_q75": 681.593681,
    "steady_tpot_ms": 433.45848200000006,
    "steady_tpot_ms_q25": 432.96234975000004,
    "steady_tpot_ms_q75": 433.5642295,
    "steady_tpot_p95_ms": 434.4063293,
    "steady_tpot_p95_ms_q25": 433.880979275,
    "steady_tpot_p95_ms_q75": 434.483423825,
    "decode_tokens_per_s": 2.2639027953876485,
    "decode_tokens_per_s_q25": 2.225618068814971,
    "decode_tokens_per_s_q75": 2.267887124487505,
    "decode_ms": 6625.726171,
    "decode_ms_q25": 6614.106202999999,
    "decode_ms_q75": 6741.695744500001,
    "end_to_end_ms": 10557.350425,
    "end_to_end_ms_q25": 10549.4645045,
    "end_to_end_ms_q75": 10589.636111,
    "output_tokens_per_s_e2e": 1.5155317722628308,
    "output_tokens_per_s_e2e_q25": 1.510925262832859,
    "output_tokens_per_s_e2e_q75": 1.5166655078770992,
    "all_tokens_per_s_e2e": 775.9522673985693,
    "all_tokens_per_s_e2e_q25": 773.5937345704239,
    "all_tokens_per_s_e2e_q75": 776.5327400330748,
    "energy_net_psys_j": 199.4748840350694,
    "energy_net_psys_j_q25": 197.87124064255613,
    "energy_net_psys_j_q75": 200.5593455579097,
    "prefill_energy_net_psys_j": 114.36186444242055,
    "prefill_energy_net_psys_j_q25": 110.84935359375453,
    "prefill_energy_net_psys_j_q75": 114.86505231196605,
    "decode_energy_net_psys_j": 86.27556689923847,
    "decode_energy_net_psys_j_q25": 85.69429324594365,
    "decode_energy_net_psys_j_q75": 87.6031607020964,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.4111026769294721,
    "energy_reduction_vs_gpu": -0.5453402330264172
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 8192,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 8102.135766,
    "ttft_ms_q25": 8100.4667055,
    "ttft_ms_q75": 8123.1352045,
    "prefill_tokens_per_s": 1009.1166374068879,
    "prefill_tokens_per_s_q25": 1006.5146565095354,
    "prefill_tokens_per_s_q75": 1009.324603674625,
    "handoff_plus_first_decode_ms": 569.7060819999999,
    "handoff_plus_first_decode_ms_q25": 551.0542674999999,
    "handoff_plus_first_decode_ms_q75": 681.593681,
    "steady_tpot_ms": 433.45848200000006,
    "steady_tpot_ms_q25": 432.96234975000004,
    "steady_tpot_ms_q75": 433.5642295,
    "steady_tpot_p95_ms": 434.4063293,
    "steady_tpot_p95_ms_q25": 433.880979275,
    "steady_tpot_p95_ms_q75": 434.483423825,
    "decode_tokens_per_s": 2.2639027953876485,
    "decode_tokens_per_s_q25": 2.225618068814971,
    "decode_tokens_per_s_q75": 2.267887124487505,
    "decode_ms": 6625.726171,
    "decode_ms_q25": 6614.106202999999,
    "decode_ms_q75": 6741.695744500001,
    "end_to_end_ms": 14769.860814,
    "end_to_end_ms_q25": 14737.2414075,
    "end_to_end_ms_q75": 14863.161888499999,
    "output_tokens_per_s_e2e": 1.0832871210833606,
    "output_tokens_per_s_e2e_q25": 1.0765293834659087,
    "output_tokens_per_s_e2e_q75": 1.0856901874768,
    "all_tokens_per_s_e2e": 554.6430059946806,
    "all_tokens_per_s_e2e_q25": 551.1830443345452,
    "all_tokens_per_s_e2e_q75": 555.8733759881216,
    "energy_net_psys_j": 222.90180490131218,
    "energy_net_psys_j_q25": 219.5411328557679,
    "energy_net_psys_j_q75": 222.93832153994018,
    "prefill_energy_net_psys_j": 134.04408367361384,
    "prefill_energy_net_psys_j_q25": 132.55576244559433,
    "prefill_energy_net_psys_j_q75": 135.33516083784377,
    "decode_energy_net_psys_j": 86.27556689923847,
    "decode_energy_net_psys_j_q25": 85.69429324594365,
    "decode_energy_net_psys_j_q75": 87.6031607020964,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.29385212735966143,
    "energy_reduction_vs_gpu": -0.7268295645061955
  },
  {
    "precision": "INT8",
    "model": "llama32_1b",
    "model_label": "Llama 3.2 1B",
    "family": "Llama 3.2",
    "size_b": 1.0,
    "context_length": 8192,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 8102.135766,
    "ttft_ms_q25": 8100.4667055,
    "ttft_ms_q75": 8123.1352045,
    "prefill_tokens_per_s": 1009.1166374068879,
    "prefill_tokens_per_s_q25": 1006.5146565095354,
    "prefill_tokens_per_s_q75": 1009.324603674625,
    "handoff_plus_first_decode_ms": 52.534369,
    "handoff_plus_first_decode_ms_q25": 50.268923,
    "handoff_plus_first_decode_ms_q75": 56.2366885,
    "steady_tpot_ms": 25.031965,
    "steady_tpot_ms_q25": 25.030341999999997,
    "steady_tpot_ms_q75": 25.149005000000002,
    "steady_tpot_p95_ms": 27.1545193,
    "steady_tpot_p95_ms_q25": 26.919350299999998,
    "steady_tpot_p95_ms_q75": 28.1688526,
    "decode_tokens_per_s": 36.71694083202306,
    "decode_tokens_per_s_q25": 36.34434649602082,
    "decode_tokens_per_s_q75": 36.752512545544135,
    "decode_ms": 408.53076700000065,
    "decode_ms_q25": 408.13574400000016,
    "decode_ms_q75": 412.76231650000045,
    "end_to_end_ms": 8515.791511,
    "end_to_end_ms_q25": 8512.833998999999,
    "end_to_end_ms_q75": 8534.228460499999,
    "output_tokens_per_s_e2e": 1.878862344073656,
    "output_tokens_per_s_e2e_q25": 1.8748120873999137,
    "output_tokens_per_s_e2e_q75": 1.8795153215801375,
    "all_tokens_per_s_e2e": 961.9775201657119,
    "all_tokens_per_s_e2e_q25": 959.9037887487558,
    "all_tokens_per_s_e2e_q75": 962.3118446490304,
    "energy_net_psys_j": 148.94974348966838,
    "energy_net_psys_j_q25": 147.36842590423583,
    "energy_net_psys_j_q75": 150.16560880242275,
    "prefill_energy_net_psys_j": 134.04408367361384,
    "prefill_energy_net_psys_j_q25": 132.55576244559433,
    "prefill_energy_net_psys_j_q75": 135.33516083784377,
    "decode_energy_net_psys_j": 14.755236113103393,
    "decode_energy_net_psys_j_q25": 14.73745160716593,
    "decode_energy_net_psys_j_q75": 14.830447964578973,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.5096596147749443,
    "energy_reduction_vs_gpu": -0.15391986528530532
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 497.33266599999996,
    "ttft_ms_q25": 478.1142695,
    "ttft_ms_q75": 501.8957445,
    "prefill_tokens_per_s": 2026.8123710981013,
    "prefill_tokens_per_s_q25": 2008.551253838245,
    "prefill_tokens_per_s_q75": 2111.694554771321,
    "handoff_plus_first_decode_ms": 53.068780000000004,
    "handoff_plus_first_decode_ms_q25": 52.72070050000001,
    "handoff_plus_first_decode_ms_q75": 59.843194,
    "steady_tpot_ms": 44.791843,
    "steady_tpot_ms_q25": 44.789614,
    "steady_tpot_ms_q75": 44.807373999999996,
    "steady_tpot_p95_ms": 46.701485,
    "steady_tpot_p95_ms_q25": 46.1616997,
    "steady_tpot_p95_ms_q75": 48.984910400000004,
    "decode_tokens_per_s": 21.909214204853633,
    "decode_tokens_per_s_q25": 21.562931642814615,
    "decode_tokens_per_s_q75": 21.967048969526545,
    "decode_ms": 684.64345,
    "decode_ms_q25": 682.8456565,
    "decode_ms_q75": 695.8176975,
    "end_to_end_ms": 1181.976116,
    "end_to_end_ms_q25": 1173.931967,
    "end_to_end_ms_q75": 1184.741401,
    "output_tokens_per_s_e2e": 13.53665254603165,
    "output_tokens_per_s_e2e_q25": 13.505130447944067,
    "output_tokens_per_s_e2e_q75": 13.630049910992803,
    "all_tokens_per_s_e2e": 866.3457629460256,
    "all_tokens_per_s_e2e_q25": 864.3283486684203,
    "all_tokens_per_s_e2e_q75": 872.3231943035394,
    "energy_net_psys_j": 41.03450894235209,
    "energy_net_psys_j_q25": 40.67057957181344,
    "energy_net_psys_j_q75": 41.47673791445965,
    "prefill_energy_net_psys_j": 16.793194167137834,
    "prefill_energy_net_psys_j_q25": 16.648150185324507,
    "prefill_energy_net_psys_j_q75": 16.88160164530715,
    "decode_energy_net_psys_j": 24.064499818875625,
    "decode_energy_net_psys_j_q25": 23.78897792650629,
    "decode_energy_net_psys_j_q75": 24.74018025096582,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 497.33266599999996,
    "ttft_ms_q25": 478.1142695,
    "ttft_ms_q75": 501.8957445,
    "prefill_tokens_per_s": 2026.8123710981013,
    "prefill_tokens_per_s_q25": 2008.551253838245,
    "prefill_tokens_per_s_q75": 2111.694554771321,
    "handoff_plus_first_decode_ms": 2837.053767,
    "handoff_plus_first_decode_ms_q25": 2836.1596645,
    "handoff_plus_first_decode_ms_q75": 2958.2273855000003,
    "steady_tpot_ms": 2828.334214,
    "steady_tpot_ms_q25": 2826.94664625,
    "steady_tpot_ms_q75": 2828.9700942500003,
    "steady_tpot_p95_ms": 2833.54747725,
    "steady_tpot_p95_ms_q25": 2833.211803225,
    "steady_tpot_p95_ms_q75": 2835.318217425,
    "decode_tokens_per_s": 0.3533523389066604,
    "decode_tokens_per_s_q25": 0.3528598698968898,
    "decode_tokens_per_s_q75": 0.35521912487418994,
    "decode_ms": 42450.546801000004,
    "decode_ms_q25": 42228.622377,
    "decode_ms_q75": 42509.875725000005,
    "end_to_end_ms": 42947.879467000006,
    "end_to_end_ms_q25": 42730.518121500005,
    "end_to_end_ms_q75": 42987.98999450001,
    "output_tokens_per_s_e2e": 0.37254458656786466,
    "output_tokens_per_s_e2e_q25": 0.37219730282298114,
    "output_tokens_per_s_e2e_q75": 0.37444933320951723,
    "all_tokens_per_s_e2e": 23.842853540343338,
    "all_tokens_per_s_e2e_q25": 23.820627380670793,
    "all_tokens_per_s_e2e_q75": 23.964757325409103,
    "energy_net_psys_j": 321.90318959507783,
    "energy_net_psys_j_q25": 313.6633753237112,
    "energy_net_psys_j_q75": 329.03898013965096,
    "prefill_energy_net_psys_j": 16.793194167137834,
    "prefill_energy_net_psys_j_q25": 16.648150185324507,
    "prefill_energy_net_psys_j_q75": 16.88160164530715,
    "decode_energy_net_psys_j": 305.4000833915667,
    "decode_energy_net_psys_j_q25": 296.92681766021735,
    "decode_energy_net_psys_j_q75": 312.3908299543265,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.027521175216769403,
    "energy_reduction_vs_gpu": -6.844694572738962
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3434.224864,
    "ttft_ms_q25": 3431.2598755,
    "ttft_ms_q75": 3434.5322255,
    "prefill_tokens_per_s": 293.5160159623141,
    "prefill_tokens_per_s_q25": 293.48975111918094,
    "prefill_tokens_per_s_q75": 293.769865657448,
    "handoff_plus_first_decode_ms": 2837.053767,
    "handoff_plus_first_decode_ms_q25": 2836.1596645,
    "handoff_plus_first_decode_ms_q75": 2958.2273855000003,
    "steady_tpot_ms": 2828.334214,
    "steady_tpot_ms_q25": 2826.94664625,
    "steady_tpot_ms_q75": 2828.9700942500003,
    "steady_tpot_p95_ms": 2833.54747725,
    "steady_tpot_p95_ms_q25": 2833.211803225,
    "steady_tpot_p95_ms_q75": 2835.318217425,
    "decode_tokens_per_s": 0.3533523389066604,
    "decode_tokens_per_s_q25": 0.3528598698968898,
    "decode_tokens_per_s_q75": 0.35521912487418994,
    "decode_ms": 42450.546801000004,
    "decode_ms_q25": 42228.622377,
    "decode_ms_q75": 42509.875725000005,
    "end_to_end_ms": 45884.77166500001,
    "end_to_end_ms_q25": 45659.8822525,
    "end_to_end_ms_q75": 45944.407950500005,
    "output_tokens_per_s_e2e": 0.3486995667498217,
    "output_tokens_per_s_e2e_q25": 0.3482475380801305,
    "output_tokens_per_s_e2e_q75": 0.3504255236786921,
    "all_tokens_per_s_e2e": 22.31677227198859,
    "all_tokens_per_s_e2e_q25": 22.28784243712835,
    "all_tokens_per_s_e2e_q75": 22.427233515436296,
    "energy_net_psys_j": 344.8710458081819,
    "energy_net_psys_j_q25": 336.6444824276074,
    "energy_net_psys_j_q75": 352.4744276629308,
    "prefill_energy_net_psys_j": 39.96436711816477,
    "prefill_energy_net_psys_j_q25": 39.71766476739,
    "prefill_energy_net_psys_j_q75": 40.330300059379056,
    "decode_energy_net_psys_j": 305.4000833915667,
    "decode_energy_net_psys_j_q25": 296.92681766021735,
    "decode_energy_net_psys_j_q75": 312.3908299543265,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.025759659972364818,
    "energy_reduction_vs_gpu": -7.404415081283874
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3434.224864,
    "ttft_ms_q25": 3431.2598755,
    "ttft_ms_q75": 3434.5322255,
    "prefill_tokens_per_s": 293.5160159623141,
    "prefill_tokens_per_s_q25": 293.48975111918094,
    "prefill_tokens_per_s_q75": 293.769865657448,
    "handoff_plus_first_decode_ms": 53.068780000000004,
    "handoff_plus_first_decode_ms_q25": 52.72070050000001,
    "handoff_plus_first_decode_ms_q75": 59.843194,
    "steady_tpot_ms": 44.791843,
    "steady_tpot_ms_q25": 44.789614,
    "steady_tpot_ms_q75": 44.807373999999996,
    "steady_tpot_p95_ms": 46.701485,
    "steady_tpot_p95_ms_q25": 46.1616997,
    "steady_tpot_p95_ms_q75": 48.984910400000004,
    "decode_tokens_per_s": 21.909214204853633,
    "decode_tokens_per_s_q25": 21.562931642814615,
    "decode_tokens_per_s_q75": 21.967048969526545,
    "decode_ms": 684.64345,
    "decode_ms_q25": 682.8456565,
    "decode_ms_q75": 695.8176975,
    "end_to_end_ms": 4118.868313999999,
    "end_to_end_ms_q25": 4114.105532,
    "end_to_end_ms_q75": 4130.349923,
    "output_tokens_per_s_e2e": 3.8845621613140997,
    "output_tokens_per_s_e2e_q25": 3.8737937310851502,
    "output_tokens_per_s_e2e_q75": 3.889064419671543,
    "all_tokens_per_s_e2e": 248.61197832410238,
    "all_tokens_per_s_e2e_q25": 247.92279878944962,
    "all_tokens_per_s_e2e_q75": 248.90012285897876,
    "energy_net_psys_j": 64.2096890347303,
    "energy_net_psys_j_q25": 64.11927798588535,
    "energy_net_psys_j_q75": 64.54825606720077,
    "prefill_energy_net_psys_j": 39.96436711816477,
    "prefill_energy_net_psys_j_q25": 39.71766476739,
    "prefill_energy_net_psys_j_q75": 40.330300059379056,
    "decode_energy_net_psys_j": 24.064499818875625,
    "decode_energy_net_psys_j_q25": 23.78897792650629,
    "decode_energy_net_psys_j_q75": 24.74018025096582,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.2869662309869128,
    "energy_reduction_vs_gpu": -0.564772935992391
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1086.387235,
    "ttft_ms_q25": 1065.4610715,
    "ttft_ms_q75": 1086.429896,
    "prefill_tokens_per_s": 1870.4196206797296,
    "prefill_tokens_per_s_q25": 1870.3461775256214,
    "prefill_tokens_per_s_q75": 1907.8915198916916,
    "handoff_plus_first_decode_ms": 62.059429,
    "handoff_plus_first_decode_ms_q25": 59.75551,
    "handoff_plus_first_decode_ms_q75": 66.2599335,
    "steady_tpot_ms": 46.5672985,
    "steady_tpot_ms_q25": 46.55176075,
    "steady_tpot_ms_q75": 46.6040225,
    "steady_tpot_p95_ms": 47.932761899999996,
    "steady_tpot_p95_ms_q25": 47.827453725,
    "steady_tpot_p95_ms_q75": 49.535231825,
    "decode_tokens_per_s": 20.920751040234133,
    "decode_tokens_per_s_q25": 20.66714167113541,
    "decode_tokens_per_s_q75": 20.975931423823205,
    "decode_ms": 716.9914679999999,
    "decode_ms_q25": 715.1102615,
    "decode_ms_q75": 725.8990765,
    "end_to_end_ms": 1799.701612,
    "end_to_end_ms_q25": 1789.5216025,
    "end_to_end_ms_q75": 1801.5401575,
    "output_tokens_per_s_e2e": 8.890362654184253,
    "output_tokens_per_s_e2e_q25": 8.881298924960864,
    "output_tokens_per_s_e2e_q75": 8.941226398748022,
    "all_tokens_per_s_e2e": 1137.9664197355844,
    "all_tokens_per_s_e2e_q25": 1136.8062623949907,
    "all_tokens_per_s_e2e_q75": 1144.4769790397468,
    "energy_net_psys_j": 62.07486891919182,
    "energy_net_psys_j_q25": 61.8576837706784,
    "energy_net_psys_j_q75": 62.25892217557576,
    "prefill_energy_net_psys_j": 37.81409887432489,
    "prefill_energy_net_psys_j_q25": 37.07612097746711,
    "prefill_energy_net_psys_j_q75": 37.83269025366914,
    "decode_energy_net_psys_j": 24.591693798946324,
    "decode_energy_net_psys_j_q25": 24.426231921906627,
    "decode_energy_net_psys_j_q75": 24.947024670250986,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1086.387235,
    "ttft_ms_q25": 1065.4610715,
    "ttft_ms_q75": 1086.429896,
    "prefill_tokens_per_s": 1870.4196206797296,
    "prefill_tokens_per_s_q25": 1870.3461775256214,
    "prefill_tokens_per_s_q75": 1907.8915198916916,
    "handoff_plus_first_decode_ms": 2911.1904329999998,
    "handoff_plus_first_decode_ms_q25": 2909.095489,
    "handoff_plus_first_decode_ms_q75": 2997.206625,
    "steady_tpot_ms": 2830.3012484999995,
    "steady_tpot_ms_q25": 2816.40385725,
    "steady_tpot_ms_q75": 2831.04298025,
    "steady_tpot_p95_ms": 2834.8560459999994,
    "steady_tpot_p95_ms_q25": 2829.7039638999995,
    "steady_tpot_p95_ms_q75": 2835.052186,
    "decode_tokens_per_s": 0.35320610810309805,
    "decode_tokens_per_s_q25": 0.3523833832510136,
    "decode_tokens_per_s_q75": 0.3552997912762771,
    "decode_ms": 42468.12174500001,
    "decode_ms_q25": 42219.33490850001,
    "decode_ms_q75": 42567.505957500005,
    "end_to_end_ms": 43554.50898000001,
    "end_to_end_ms_q25": 43305.76480450001,
    "end_to_end_ms_q75": 43632.96702900001,
    "output_tokens_per_s_e2e": 0.3673557657910254,
    "output_tokens_per_s_e2e_q25": 0.366696395463441,
    "output_tokens_per_s_e2e_q75": 0.36947801219223875,
    "all_tokens_per_s_e2e": 47.02153802125125,
    "all_tokens_per_s_e2e_q25": 46.93713861932045,
    "all_tokens_per_s_e2e_q75": 47.29318556060656,
    "energy_net_psys_j": 354.8273616203003,
    "energy_net_psys_j_q25": 339.1192137041812,
    "energy_net_psys_j_q75": 356.4453942628092,
    "prefill_energy_net_psys_j": 37.81409887432489,
    "prefill_energy_net_psys_j_q25": 37.07612097746711,
    "prefill_energy_net_psys_j_q75": 37.83269025366914,
    "decode_energy_net_psys_j": 316.9760799872869,
    "decode_energy_net_psys_j_q25": 301.28652345051205,
    "decode_energy_net_psys_j_q75": 319.3506819059978,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.041320672741975185,
    "energy_reduction_vs_gpu": -4.71611938612725
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 6953.09225,
    "ttft_ms_q25": 6939.7785735,
    "ttft_ms_q75": 6960.2044145,
    "prefill_tokens_per_s": 292.24407313163437,
    "prefill_tokens_per_s_q25": 291.9457534128883,
    "prefill_tokens_per_s_q75": 292.8058088953429,
    "handoff_plus_first_decode_ms": 2911.1904329999998,
    "handoff_plus_first_decode_ms_q25": 2909.095489,
    "handoff_plus_first_decode_ms_q75": 2997.206625,
    "steady_tpot_ms": 2830.3012484999995,
    "steady_tpot_ms_q25": 2816.40385725,
    "steady_tpot_ms_q75": 2831.04298025,
    "steady_tpot_p95_ms": 2834.8560459999994,
    "steady_tpot_p95_ms_q25": 2829.7039638999995,
    "steady_tpot_p95_ms_q75": 2835.052186,
    "decode_tokens_per_s": 0.35320610810309805,
    "decode_tokens_per_s_q25": 0.3523833832510136,
    "decode_tokens_per_s_q75": 0.3552997912762771,
    "decode_ms": 42468.12174500001,
    "decode_ms_q25": 42219.33490850001,
    "decode_ms_q75": 42567.505957500005,
    "end_to_end_ms": 49394.58664200001,
    "end_to_end_ms_q25": 49166.2256465,
    "end_to_end_ms_q75": 49507.284531000005,
    "output_tokens_per_s_e2e": 0.32392213575880535,
    "output_tokens_per_s_e2e_q25": 0.3231864373744046,
    "output_tokens_per_s_e2e_q75": 0.32543366843220445,
    "all_tokens_per_s_e2e": 41.462033377127085,
    "all_tokens_per_s_e2e_q25": 41.36786398392379,
    "all_tokens_per_s_e2e_q75": 41.65550955932217,
    "energy_net_psys_j": 397.6464616747197,
    "energy_net_psys_j_q25": 384.4032463141042,
    "energy_net_psys_j_q75": 400.91593869045266,
    "prefill_energy_net_psys_j": 82.46013188147681,
    "prefill_energy_net_psys_j_q25": 81.56525678445485,
    "prefill_energy_net_psys_j_q75": 84.01159796061407,
    "decode_energy_net_psys_j": 316.9760799872869,
    "decode_energy_net_psys_j_q25": 301.28652345051205,
    "decode_energy_net_psys_j_q75": 319.3506819059978,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.036435199367975304,
    "energy_reduction_vs_gpu": -5.4059170578736175
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 6953.09225,
    "ttft_ms_q25": 6939.7785735,
    "ttft_ms_q75": 6960.2044145,
    "prefill_tokens_per_s": 292.24407313163437,
    "prefill_tokens_per_s_q25": 291.9457534128883,
    "prefill_tokens_per_s_q75": 292.8058088953429,
    "handoff_plus_first_decode_ms": 62.059429,
    "handoff_plus_first_decode_ms_q25": 59.75551,
    "handoff_plus_first_decode_ms_q75": 66.2599335,
    "steady_tpot_ms": 46.5672985,
    "steady_tpot_ms_q25": 46.55176075,
    "steady_tpot_ms_q75": 46.6040225,
    "steady_tpot_p95_ms": 47.932761899999996,
    "steady_tpot_p95_ms_q25": 47.827453725,
    "steady_tpot_p95_ms_q75": 49.535231825,
    "decode_tokens_per_s": 20.920751040234133,
    "decode_tokens_per_s_q25": 20.66714167113541,
    "decode_tokens_per_s_q75": 20.975931423823205,
    "decode_ms": 716.9914679999999,
    "decode_ms_q25": 715.1102615,
    "decode_ms_q75": 725.8990765,
    "end_to_end_ms": 7680.545634,
    "end_to_end_ms_q25": 7662.0009995,
    "end_to_end_ms_q75": 7684.2222845,
    "output_tokens_per_s_e2e": 2.0831853311529973,
    "output_tokens_per_s_e2e_q25": 2.0821890713876012,
    "output_tokens_per_s_e2e_q75": 2.088239577388888,
    "all_tokens_per_s_e2e": 266.64772238758366,
    "all_tokens_per_s_e2e_q25": 266.52020113761296,
    "all_tokens_per_s_e2e_q75": 267.29466590577766,
    "energy_net_psys_j": 107.76248742303247,
    "energy_net_psys_j_q25": 106.51228145470583,
    "energy_net_psys_j_q75": 108.79316075382536,
    "prefill_energy_net_psys_j": 82.46013188147681,
    "prefill_energy_net_psys_j_q25": 81.56525678445485,
    "prefill_energy_net_psys_j_q75": 84.01159796061407,
    "decode_energy_net_psys_j": 24.591693798946324,
    "decode_energy_net_psys_j_q25": 24.426231921906627,
    "decode_energy_net_psys_j_q75": 24.947024670250986,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.23431949991067524,
    "energy_reduction_vs_gpu": -0.7360082960999264
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2720.652013,
    "ttft_ms_q25": 2673.5443784999998,
    "ttft_ms_q75": 2721.772441,
    "prefill_tokens_per_s": 1499.6405201784987,
    "prefill_tokens_per_s_q25": 1499.0234414649467,
    "prefill_tokens_per_s_q75": 1526.5379955002797,
    "handoff_plus_first_decode_ms": 67.901695,
    "handoff_plus_first_decode_ms_q25": 67.59164899999999,
    "handoff_plus_first_decode_ms_q75": 72.556673,
    "steady_tpot_ms": 49.9552465,
    "steady_tpot_ms_q25": 49.91485,
    "steady_tpot_ms_q75": 50.310657000000006,
    "steady_tpot_p95_ms": 51.7227957,
    "steady_tpot_p95_ms_q25": 51.69764555,
    "steady_tpot_p95_ms_q75": 53.02946695,
    "decode_tokens_per_s": 19.294544304366095,
    "decode_tokens_per_s_q25": 19.155190280478728,
    "decode_tokens_per_s_q75": 19.37388034520471,
    "decode_ms": 777.4218330000003,
    "decode_ms_q25": 774.2512740000002,
    "decode_ms_q75": 783.1190245000005,
    "end_to_end_ms": 3493.973584,
    "end_to_end_ms_q25": 3454.613272,
    "end_to_end_ms_q75": 3496.0237150000003,
    "output_tokens_per_s_e2e": 4.579313384986371,
    "output_tokens_per_s_e2e_q25": 4.576629567336262,
    "output_tokens_per_s_e2e_q75": 4.632089330337967,
    "all_tokens_per_s_e2e": 1172.304226556511,
    "all_tokens_per_s_e2e_q25": 1171.617169238083,
    "all_tokens_per_s_e2e_q75": 1185.8148685665196,
    "energy_net_psys_j": 120.78508402481626,
    "energy_net_psys_j_q25": 118.99958443530213,
    "energy_net_psys_j_q75": 121.30392492664456,
    "prefill_energy_net_psys_j": 94.75144981319067,
    "prefill_energy_net_psys_j_q25": 92.4771215125366,
    "prefill_energy_net_psys_j_q75": 94.77767278372785,
    "decode_energy_net_psys_j": 27.011291633905472,
    "decode_energy_net_psys_j_q25": 26.522462922765527,
    "decode_energy_net_psys_j_q75": 27.01508085405664,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2720.652013,
    "ttft_ms_q25": 2673.5443784999998,
    "ttft_ms_q75": 2721.772441,
    "prefill_tokens_per_s": 1499.6405201784987,
    "prefill_tokens_per_s_q25": 1499.0234414649467,
    "prefill_tokens_per_s_q75": 1526.5379955002797,
    "handoff_plus_first_decode_ms": 3007.338554,
    "handoff_plus_first_decode_ms_q25": 3000.3197135,
    "handoff_plus_first_decode_ms_q75": 3082.431552,
    "steady_tpot_ms": 2855.4519525,
    "steady_tpot_ms_q25": 2853.6767274999997,
    "steady_tpot_ms_q75": 2856.394367,
    "steady_tpot_p95_ms": 2862.3974482999997,
    "steady_tpot_p95_ms_q25": 2861.166444,
    "steady_tpot_p95_ms_q75": 2862.480444275,
    "decode_tokens_per_s": 0.3491382514618317,
    "decode_tokens_per_s_q25": 0.3486594938836961,
    "decode_tokens_per_s_q75": 0.34977202156965304,
    "decode_ms": 42962.923533,
    "decode_ms_q25": 42885.217577999996,
    "decode_ms_q75": 43021.9986715,
    "end_to_end_ms": 45683.575546,
    "end_to_end_ms_q25": 45606.990019000004,
    "end_to_end_ms_q75": 45695.54305,
    "output_tokens_per_s_e2e": 0.3502352827853673,
    "output_tokens_per_s_e2e_q25": 0.35014358139763313,
    "output_tokens_per_s_e2e_q75": 0.35082440458222774,
    "all_tokens_per_s_e2e": 89.66023239305403,
    "all_tokens_per_s_e2e_q25": 89.63675683779408,
    "all_tokens_per_s_e2e_q75": 89.8110475730503,
    "energy_net_psys_j": 413.19929953653985,
    "energy_net_psys_j_q25": 411.94730694220914,
    "energy_net_psys_j_q75": 419.2861992388679,
    "prefill_energy_net_psys_j": 94.75144981319067,
    "prefill_energy_net_psys_j_q25": 92.4771215125366,
    "prefill_energy_net_psys_j_q75": 94.77767278372785,
    "decode_energy_net_psys_j": 322.99650632465733,
    "decode_energy_net_psys_j_q25": 319.4439624591353,
    "decode_energy_net_psys_j_q75": 326.8090777263312,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.0764820516398027,
    "energy_reduction_vs_gpu": -2.4209464096712865
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 14660.231373,
    "ttft_ms_q25": 14638.466973999999,
    "ttft_ms_q75": 14668.88678,
    "prefill_tokens_per_s": 278.30392960333535,
    "prefill_tokens_per_s_q25": 278.139812631316,
    "prefill_tokens_per_s_q75": 278.71832660813686,
    "handoff_plus_first_decode_ms": 3007.338554,
    "handoff_plus_first_decode_ms_q25": 3000.3197135,
    "handoff_plus_first_decode_ms_q75": 3082.431552,
    "steady_tpot_ms": 2855.4519525,
    "steady_tpot_ms_q25": 2853.6767274999997,
    "steady_tpot_ms_q75": 2856.394367,
    "steady_tpot_p95_ms": 2862.3974482999997,
    "steady_tpot_p95_ms_q25": 2861.166444,
    "steady_tpot_p95_ms_q75": 2862.480444275,
    "decode_tokens_per_s": 0.3491382514618317,
    "decode_tokens_per_s_q25": 0.3486594938836961,
    "decode_tokens_per_s_q75": 0.34977202156965304,
    "decode_ms": 42962.923533,
    "decode_ms_q25": 42885.217577999996,
    "decode_ms_q75": 43021.9986715,
    "end_to_end_ms": 57640.46572,
    "end_to_end_ms_q25": 57532.339959000004,
    "end_to_end_ms_q75": 57690.8854515,
    "output_tokens_per_s_e2e": 0.2775827675946127,
    "output_tokens_per_s_e2e_q25": 0.2773403822074658,
    "output_tokens_per_s_e2e_q75": 0.27810543644443214,
    "all_tokens_per_s_e2e": 71.06118850422085,
    "all_tokens_per_s_e2e_q25": 70.99913784511125,
    "all_tokens_per_s_e2e_q75": 71.19499172977463,
    "energy_net_psys_j": 502.74846344553436,
    "energy_net_psys_j_q25": 496.5750959106758,
    "energy_net_psys_j_q75": 509.4081391164607,
    "prefill_energy_net_psys_j": 179.75195712087702,
    "prefill_energy_net_psys_j_q25": 177.13113345154053,
    "prefill_energy_net_psys_j_q75": 182.5990613901294,
    "decode_energy_net_psys_j": 322.99650632465733,
    "decode_energy_net_psys_j_q25": 319.4439624591353,
    "decode_energy_net_psys_j_q75": 326.8090777263312,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.06061667858432425,
    "energy_reduction_vs_gpu": -3.162338980053536
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 14660.231373,
    "ttft_ms_q25": 14638.466973999999,
    "ttft_ms_q75": 14668.88678,
    "prefill_tokens_per_s": 278.30392960333535,
    "prefill_tokens_per_s_q25": 278.139812631316,
    "prefill_tokens_per_s_q75": 278.71832660813686,
    "handoff_plus_first_decode_ms": 67.901695,
    "handoff_plus_first_decode_ms_q25": 67.59164899999999,
    "handoff_plus_first_decode_ms_q75": 72.556673,
    "steady_tpot_ms": 49.9552465,
    "steady_tpot_ms_q25": 49.91485,
    "steady_tpot_ms_q75": 50.310657000000006,
    "steady_tpot_p95_ms": 51.7227957,
    "steady_tpot_p95_ms_q25": 51.69764555,
    "steady_tpot_p95_ms_q75": 53.02946695,
    "decode_tokens_per_s": 19.294544304366095,
    "decode_tokens_per_s_q25": 19.155190280478728,
    "decode_tokens_per_s_q75": 19.37388034520471,
    "decode_ms": 777.4218330000003,
    "decode_ms_q25": 774.2512740000002,
    "decode_ms_q75": 783.1190245000005,
    "end_to_end_ms": 15449.047589000002,
    "end_to_end_ms_q25": 15418.4154395,
    "end_to_end_ms_q75": 15452.0058045,
    "output_tokens_per_s_e2e": 1.0356625486345377,
    "output_tokens_per_s_e2e_q25": 1.0354643137502522,
    "output_tokens_per_s_e2e_q75": 1.0377242211589968,
    "all_tokens_per_s_e2e": 265.12961245044164,
    "all_tokens_per_s_e2e_q25": 265.07886432006455,
    "all_tokens_per_s_e2e_q75": 265.65740061670317,
    "energy_net_psys_j": 206.7632487547825,
    "energy_net_psys_j_q25": 204.1462143055972,
    "energy_net_psys_j_q75": 209.12152431289493,
    "prefill_energy_net_psys_j": 179.75195712087702,
    "prefill_energy_net_psys_j_q25": 177.13113345154053,
    "prefill_energy_net_psys_j_q75": 182.5990613901294,
    "decode_energy_net_psys_j": 27.011291633905472,
    "decode_energy_net_psys_j_q25": 26.522462922765527,
    "decode_energy_net_psys_j_q75": 27.01508085405664,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.22616109917919935,
    "energy_reduction_vs_gpu": -0.7118276683261762
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 7314.639501000001,
    "ttft_ms_q25": 7211.2596115,
    "ttft_ms_q75": 7314.9642140000005,
    "prefill_tokens_per_s": 1117.7584348322623,
    "prefill_tokens_per_s_q25": 1117.7088194721211,
    "prefill_tokens_per_s_q75": 1134.0155674794278,
    "handoff_plus_first_decode_ms": 82.083003,
    "handoff_plus_first_decode_ms_q25": 81.43185700000001,
    "handoff_plus_first_decode_ms_q75": 87.66125600000001,
    "steady_tpot_ms": 56.947846999999996,
    "steady_tpot_ms_q25": 56.93222374999999,
    "steady_tpot_ms_q75": 56.966063,
    "steady_tpot_p95_ms": 59.96024355,
    "steady_tpot_p95_ms_q25": 59.65237065,
    "steady_tpot_p95_ms_q75": 60.101113049999995,
    "decode_tokens_per_s": 16.93331856928576,
    "decode_tokens_per_s_q25": 16.824421064849883,
    "decode_tokens_per_s_q75": 16.945296621925877,
    "decode_ms": 885.8275440000007,
    "decode_ms_q25": 885.2018250000001,
    "decode_ms_q75": 891.5984915000004,
    "end_to_end_ms": 8199.215607,
    "end_to_end_ms_q25": 8102.232384,
    "end_to_end_ms_q75": 8200.166039,
    "output_tokens_per_s_e2e": 1.9514061791886723,
    "output_tokens_per_s_e2e_q25": 1.9511800296291266,
    "output_tokens_per_s_e2e_q75": 1.9750473747134585,
    "all_tokens_per_s_e2e": 999.1199637446002,
    "all_tokens_per_s_e2e_q25": 999.0041751701128,
    "all_tokens_per_s_e2e_q75": 1011.2242558532907,
    "energy_net_psys_j": 277.01937880869855,
    "energy_net_psys_j_q25": 272.58850604219083,
    "energy_net_psys_j_q75": 277.4082601770699,
    "prefill_energy_net_psys_j": 246.5728915708733,
    "prefill_energy_net_psys_j_q25": 242.75006325247682,
    "prefill_energy_net_psys_j_q75": 246.68237356609126,
    "decode_energy_net_psys_j": 30.446487237825238,
    "decode_energy_net_psys_j_q25": 29.83844278971398,
    "decode_energy_net_psys_j_q75": 30.725886610978648,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 7314.639501000001,
    "ttft_ms_q25": 7211.2596115,
    "ttft_ms_q75": 7314.9642140000005,
    "prefill_tokens_per_s": 1117.7584348322623,
    "prefill_tokens_per_s_q25": 1117.7088194721211,
    "prefill_tokens_per_s_q75": 1134.0155674794278,
    "handoff_plus_first_decode_ms": 3123.8754270000004,
    "handoff_plus_first_decode_ms_q25": 3108.1939005000004,
    "handoff_plus_first_decode_ms_q75": 3194.406264,
    "steady_tpot_ms": 2848.5759369999996,
    "steady_tpot_ms_q25": 2841.4786969999996,
    "steady_tpot_ms_q75": 2852.1456874999994,
    "steady_tpot_p95_ms": 2858.8260909,
    "steady_tpot_p95_ms_q25": 2857.403378,
    "steady_tpot_p95_ms_q75": 2860.3973753,
    "decode_tokens_per_s": 0.3491939275875777,
    "decode_tokens_per_s_q25": 0.34817720064221264,
    "decode_tokens_per_s_q75": 0.3500286445547708,
    "decode_ms": 42956.07344500001,
    "decode_ms_q25": 42853.879351,
    "decode_ms_q75": 43081.878658,
    "end_to_end_ms": 50270.71294600001,
    "end_to_end_ms_q25": 50168.843565,
    "end_to_end_ms_q75": 50293.13826950001,
    "output_tokens_per_s_e2e": 0.3182767671742977,
    "output_tokens_per_s_e2e_q25": 0.31813491326356447,
    "output_tokens_per_s_e2e_q75": 0.3189243528889761,
    "all_tokens_per_s_e2e": 162.95770479324042,
    "all_tokens_per_s_e2e_q25": 162.885075590945,
    "all_tokens_per_s_e2e_q75": 163.28926867915575,
    "energy_net_psys_j": 564.3997199908467,
    "energy_net_psys_j_q25": 560.1001275758556,
    "energy_net_psys_j_q75": 566.0037019722143,
    "prefill_energy_net_psys_j": 246.5728915708733,
    "prefill_energy_net_psys_j_q25": 242.75006325247682,
    "prefill_energy_net_psys_j_q75": 246.68237356609126,
    "decode_energy_net_psys_j": 317.8268284199735,
    "decode_energy_net_psys_j_q25": 313.4177540097644,
    "decode_energy_net_psys_j_q75": 323.2536387197376,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.16310123979756297,
    "energy_reduction_vs_gpu": -1.037401579694555
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 31436.051444,
    "ttft_ms_q25": 31423.211943000002,
    "ttft_ms_q75": 31491.0653075,
    "prefill_tokens_per_s": 260.0835545317986,
    "prefill_tokens_per_s_q25": 259.62998941234116,
    "prefill_tokens_per_s_q75": 260.18986791807447,
    "handoff_plus_first_decode_ms": 3123.8754270000004,
    "handoff_plus_first_decode_ms_q25": 3108.1939005000004,
    "handoff_plus_first_decode_ms_q75": 3194.406264,
    "steady_tpot_ms": 2848.5759369999996,
    "steady_tpot_ms_q25": 2841.4786969999996,
    "steady_tpot_ms_q75": 2852.1456874999994,
    "steady_tpot_p95_ms": 2858.8260909,
    "steady_tpot_p95_ms_q25": 2857.403378,
    "steady_tpot_p95_ms_q75": 2860.3973753,
    "decode_tokens_per_s": 0.3491939275875777,
    "decode_tokens_per_s_q25": 0.34817720064221264,
    "decode_tokens_per_s_q75": 0.3500286445547708,
    "decode_ms": 42956.07344500001,
    "decode_ms_q25": 42853.879351,
    "decode_ms_q75": 43081.878658,
    "end_to_end_ms": 74502.152616,
    "end_to_end_ms_q25": 74332.1051575,
    "end_to_end_ms_q75": 74572.9439655,
    "output_tokens_per_s_e2e": 0.2147588953901428,
    "output_tokens_per_s_e2e_q25": 0.21455522026646848,
    "output_tokens_per_s_e2e_q75": 0.2152513197575947,
    "all_tokens_per_s_e2e": 109.95655443975312,
    "all_tokens_per_s_e2e_q25": 109.85227277643186,
    "all_tokens_per_s_e2e_q75": 110.20867571588849,
    "energy_net_psys_j": 702.089033212181,
    "energy_net_psys_j_q25": 699.998348089371,
    "energy_net_psys_j_q75": 710.8013081087704,
    "prefill_energy_net_psys_j": 388.8989833670057,
    "prefill_energy_net_psys_j_q25": 381.1537837798425,
    "prefill_energy_net_psys_j_q75": 395.2928689761961,
    "decode_energy_net_psys_j": 317.8268284199735,
    "decode_energy_net_psys_j_q25": 313.4177540097644,
    "decode_energy_net_psys_j_q75": 323.2536387197376,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.11005340542655871,
    "energy_reduction_vs_gpu": -1.5344401400055951
  },
  {
    "precision": "INT8",
    "model": "llama32_3b",
    "model_label": "Llama 3.2 3B",
    "family": "Llama 3.2",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 31436.051444,
    "ttft_ms_q25": 31423.211943000002,
    "ttft_ms_q75": 31491.0653075,
    "prefill_tokens_per_s": 260.0835545317986,
    "prefill_tokens_per_s_q25": 259.62998941234116,
    "prefill_tokens_per_s_q75": 260.18986791807447,
    "handoff_plus_first_decode_ms": 82.083003,
    "handoff_plus_first_decode_ms_q25": 81.43185700000001,
    "handoff_plus_first_decode_ms_q75": 87.66125600000001,
    "steady_tpot_ms": 56.947846999999996,
    "steady_tpot_ms_q25": 56.93222374999999,
    "steady_tpot_ms_q75": 56.966063,
    "steady_tpot_p95_ms": 59.96024355,
    "steady_tpot_p95_ms_q25": 59.65237065,
    "steady_tpot_p95_ms_q75": 60.101113049999995,
    "decode_tokens_per_s": 16.93331856928576,
    "decode_tokens_per_s_q25": 16.824421064849883,
    "decode_tokens_per_s_q75": 16.945296621925877,
    "decode_ms": 885.8275440000007,
    "decode_ms_q25": 885.2018250000001,
    "decode_ms_q75": 891.5984915000004,
    "end_to_end_ms": 32333.420883,
    "end_to_end_ms_q25": 32314.8104345,
    "end_to_end_ms_q75": 32382.03808,
    "output_tokens_per_s_e2e": 0.4948440209248737,
    "output_tokens_per_s_e2e_q25": 0.4941021941241479,
    "output_tokens_per_s_e2e_q75": 0.4951291711645597,
    "all_tokens_per_s_e2e": 253.36013871353532,
    "all_tokens_per_s_e2e_q25": 252.98032339156373,
    "all_tokens_per_s_e2e_q75": 253.50613563625456,
    "energy_net_psys_j": 419.90426935113777,
    "energy_net_psys_j_q25": 411.2716259427099,
    "energy_net_psys_j_q75": 426.0187555871747,
    "prefill_energy_net_psys_j": 388.8989833670057,
    "prefill_energy_net_psys_j_q25": 381.1537837798425,
    "prefill_energy_net_psys_j_q75": 395.2928689761961,
    "decode_energy_net_psys_j": 30.446487237825238,
    "decode_energy_net_psys_j_q25": 29.83844278971398,
    "decode_energy_net_psys_j_q75": 30.725886610978648,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.2535833012123662,
    "energy_reduction_vs_gpu": -0.5157938450259516
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 1024,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 200.513376,
    "ttft_ms_q25": 186.1602465,
    "ttft_ms_q75": 209.012493,
    "prefill_tokens_per_s": 5027.096047697088,
    "prefill_tokens_per_s_q25": 4830.66571701444,
    "prefill_tokens_per_s_q75": 5447.070271358232,
    "handoff_plus_first_decode_ms": 23.977131,
    "handoff_plus_first_decode_ms_q25": 23.889815,
    "handoff_plus_first_decode_ms_q75": 36.138904,
    "steady_tpot_ms": 9.6898955,
    "steady_tpot_ms_q25": 9.58769475,
    "steady_tpot_ms_q75": 9.7572455,
    "steady_tpot_p95_ms": 12.9956709,
    "steady_tpot_p95_ms_q25": 12.935077375,
    "steady_tpot_p95_ms_q75": 13.069032825,
    "decode_tokens_per_s": 85.71541079027763,
    "decode_tokens_per_s_q25": 80.40469639304747,
    "decode_tokens_per_s_q75": 85.9918562621111,
    "decode_ms": 174.99770300000003,
    "decode_ms_q25": 174.43692550000003,
    "decode_ms_q75": 187.37369900000004,
    "end_to_end_ms": 374.38952400000005,
    "end_to_end_ms_q25": 372.97316800000004,
    "end_to_end_ms_q75": 383.44941850000004,
    "output_tokens_per_s_e2e": 42.73623852787077,
    "output_tokens_per_s_e2e_q25": 41.74980126494309,
    "output_tokens_per_s_e2e_q75": 42.8991469335344,
    "all_tokens_per_s_e2e": 2735.119265783729,
    "all_tokens_per_s_e2e_q25": 2671.9872809563576,
    "all_tokens_per_s_e2e_q75": 2745.5454037462014,
    "energy_net_psys_j": 11.999113096942306,
    "energy_net_psys_j_q25": 11.875437319028297,
    "energy_net_psys_j_q75": 12.063819293163622,
    "prefill_energy_net_psys_j": 4.832090152375949,
    "prefill_energy_net_psys_j_q25": 4.39745080108195,
    "prefill_energy_net_psys_j_q75": 4.881409094011685,
    "decode_energy_net_psys_j": 7.068385061294885,
    "decode_energy_net_psys_j_q25": 6.994028225016612,
    "decode_energy_net_psys_j_q75": 7.617049550445936,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 1024,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 200.513376,
    "ttft_ms_q25": 186.1602465,
    "ttft_ms_q75": 209.012493,
    "prefill_tokens_per_s": 5027.096047697088,
    "prefill_tokens_per_s_q25": 4830.66571701444,
    "prefill_tokens_per_s_q75": 5447.070271358232,
    "handoff_plus_first_decode_ms": 125.945494,
    "handoff_plus_first_decode_ms_q25": 125.673336,
    "handoff_plus_first_decode_ms_q75": 188.992034,
    "steady_tpot_ms": 123.0124375,
    "steady_tpot_ms_q25": 122.7702555,
    "steady_tpot_ms_q75": 123.02086625000001,
    "steady_tpot_p95_ms": 123.3848382,
    "steady_tpot_p95_ms_q25": 123.28280985,
    "steady_tpot_p95_ms_q75": 131.042861875,
    "decode_tokens_per_s": 8.117983810938089,
    "decode_tokens_per_s_q25": 7.774275003566619,
    "decode_tokens_per_s_q75": 8.132856657330919,
    "decode_ms": 1847.7494350000002,
    "decode_ms_q25": 1844.3765575000002,
    "decode_ms_q75": 1933.2190665000003,
    "end_to_end_ms": 2065.261045,
    "end_to_end_ms_q25": 2053.3890505,
    "end_to_end_ms_q75": 2127.87843,
    "output_tokens_per_s_e2e": 7.747204663902426,
    "output_tokens_per_s_e2e_q25": 7.525743525382344,
    "output_tokens_per_s_e2e_q75": 7.792256832773566,
    "all_tokens_per_s_e2e": 495.8210984897553,
    "all_tokens_per_s_e2e_q25": 481.64758562447,
    "all_tokens_per_s_e2e_q75": 498.7044372975082,
    "energy_net_psys_j": 38.66509081544087,
    "energy_net_psys_j_q25": 38.021615858920086,
    "energy_net_psys_j_q75": 39.18531196540823,
    "prefill_energy_net_psys_j": 4.832090152375949,
    "prefill_energy_net_psys_j_q25": 4.39745080108195,
    "prefill_energy_net_psys_j_q75": 4.881409094011685,
    "decode_energy_net_psys_j": 33.73436277979345,
    "decode_energy_net_psys_j_q25": 33.1402067649084,
    "decode_energy_net_psys_j_q75": 34.738542222690555,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.1812795166530631,
    "energy_reduction_vs_gpu": -2.222329059078022
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 1024,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 303.182513,
    "ttft_ms_q25": 301.4690725,
    "ttft_ms_q75": 303.69026599999995,
    "prefill_tokens_per_s": 3324.730011720696,
    "prefill_tokens_per_s_q25": 3319.18052904099,
    "prefill_tokens_per_s_q75": 3343.734582262502,
    "handoff_plus_first_decode_ms": 125.945494,
    "handoff_plus_first_decode_ms_q25": 125.673336,
    "handoff_plus_first_decode_ms_q75": 188.992034,
    "steady_tpot_ms": 123.0124375,
    "steady_tpot_ms_q25": 122.7702555,
    "steady_tpot_ms_q75": 123.02086625000001,
    "steady_tpot_p95_ms": 123.3848382,
    "steady_tpot_p95_ms_q25": 123.28280985,
    "steady_tpot_p95_ms_q75": 131.042861875,
    "decode_tokens_per_s": 8.117983810938089,
    "decode_tokens_per_s_q25": 7.774275003566619,
    "decode_tokens_per_s_q75": 8.132856657330919,
    "decode_ms": 1847.7494350000002,
    "decode_ms_q25": 1844.3765575000002,
    "decode_ms_q75": 1933.2190665000003,
    "end_to_end_ms": 2147.505067,
    "end_to_end_ms_q25": 2146.353383,
    "end_to_end_ms_q75": 2234.688139,
    "output_tokens_per_s_e2e": 7.450506285580746,
    "output_tokens_per_s_e2e_q25": 7.170750017078461,
    "output_tokens_per_s_e2e_q75": 7.454506202644489,
    "all_tokens_per_s_e2e": 476.8324022771677,
    "all_tokens_per_s_e2e_q25": 458.9280010930215,
    "all_tokens_per_s_e2e_q75": 477.0883969692473,
    "energy_net_psys_j": 37.677082543824994,
    "energy_net_psys_j_q25": 37.067707074004005,
    "energy_net_psys_j_q75": 38.670625487477956,
    "prefill_energy_net_psys_j": 3.9214467655432665,
    "prefill_energy_net_psys_j_q25": 3.916863809851472,
    "prefill_energy_net_psys_j_q75": 3.9320832647874044,
    "decode_energy_net_psys_j": 33.73436277979345,
    "decode_energy_net_psys_j_q25": 33.1402067649084,
    "decode_energy_net_psys_j_q75": 34.738542222690555,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.17433696886359898,
    "energy_reduction_vs_gpu": -2.1399889508022154
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 1024,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 303.182513,
    "ttft_ms_q25": 301.4690725,
    "ttft_ms_q75": 303.69026599999995,
    "prefill_tokens_per_s": 3324.730011720696,
    "prefill_tokens_per_s_q25": 3319.18052904099,
    "prefill_tokens_per_s_q75": 3343.734582262502,
    "handoff_plus_first_decode_ms": 23.977131,
    "handoff_plus_first_decode_ms_q25": 23.889815,
    "handoff_plus_first_decode_ms_q75": 36.138904,
    "steady_tpot_ms": 9.6898955,
    "steady_tpot_ms_q25": 9.58769475,
    "steady_tpot_ms_q75": 9.7572455,
    "steady_tpot_p95_ms": 12.9956709,
    "steady_tpot_p95_ms_q25": 12.935077375,
    "steady_tpot_p95_ms_q75": 13.069032825,
    "decode_tokens_per_s": 85.71541079027763,
    "decode_tokens_per_s_q25": 80.40469639304747,
    "decode_tokens_per_s_q75": 85.9918562621111,
    "decode_ms": 174.99770300000003,
    "decode_ms_q25": 174.43692550000003,
    "decode_ms_q75": 187.37369900000004,
    "end_to_end_ms": 478.07416700000005,
    "end_to_end_ms_q25": 476.41375100000005,
    "end_to_end_ms_q75": 490.5031875,
    "output_tokens_per_s_e2e": 33.46761047643053,
    "output_tokens_per_s_e2e_q25": 32.64052163606426,
    "output_tokens_per_s_e2e_q75": 33.584661062113796,
    "all_tokens_per_s_e2e": 2141.927070491554,
    "all_tokens_per_s_e2e_q25": 2088.9933847081124,
    "all_tokens_per_s_e2e_q75": 2149.418307975283,
    "energy_net_psys_j": 11.011104825326427,
    "energy_net_psys_j_q25": 10.921528534112221,
    "energy_net_psys_j_q75": 11.54913281523334,
    "prefill_energy_net_psys_j": 3.9214467655432665,
    "prefill_energy_net_psys_j_q25": 3.916863809851472,
    "prefill_energy_net_psys_j_q75": 3.9320832647874044,
    "decode_energy_net_psys_j": 7.068385061294885,
    "decode_energy_net_psys_j_q25": 6.994028225016612,
    "decode_energy_net_psys_j_q75": 7.617049550445936,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.783120172230515,
    "energy_reduction_vs_gpu": 0.08234010827580662
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 2048,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 431.237221,
    "ttft_ms_q25": 402.07436699999994,
    "ttft_ms_q75": 438.01073199999996,
    "prefill_tokens_per_s": 4712.023686842189,
    "prefill_tokens_per_s_q25": 4640.265424769299,
    "prefill_tokens_per_s_q75": 5080.518769545417,
    "handoff_plus_first_decode_ms": 41.103348999999994,
    "handoff_plus_first_decode_ms_q25": 32.619724,
    "handoff_plus_first_decode_ms_q75": 46.111813,
    "steady_tpot_ms": 10.842264,
    "steady_tpot_ms_q25": 10.804337,
    "steady_tpot_ms_q75": 10.941298,
    "steady_tpot_p95_ms": 13.5699021,
    "steady_tpot_p95_ms_q25": 13.259286475,
    "steady_tpot_p95_ms_q75": 18.472543424999998,
    "decode_tokens_per_s": 70.3662392466489,
    "decode_tokens_per_s_q25": 67.74458303239297,
    "decode_tokens_per_s_q75": 75.50219641065138,
    "decode_ms": 213.170409,
    "decode_ms_q25": 199.59328649999998,
    "decode_ms_q75": 221.75201850000005,
    "end_to_end_ms": 630.800407,
    "end_to_end_ms_q25": 608.4411645,
    "end_to_end_ms_q75": 646.185628,
    "output_tokens_per_s_e2e": 25.364599994622388,
    "output_tokens_per_s_e2e_q25": 24.774731234407614,
    "output_tokens_per_s_e2e_q75": 26.33226888340149,
    "all_tokens_per_s_e2e": 3246.6687993116657,
    "all_tokens_per_s_e2e_q25": 3171.1655980041746,
    "all_tokens_per_s_e2e_q75": 3370.5304170753907,
    "energy_net_psys_j": 18.762792672588922,
    "energy_net_psys_j_q25": 18.227245520253454,
    "energy_net_psys_j_q75": 19.007359662895162,
    "prefill_energy_net_psys_j": 11.128946290697668,
    "prefill_energy_net_psys_j_q25": 10.176572668790886,
    "prefill_energy_net_psys_j_q75": 11.236914521531057,
    "decode_energy_net_psys_j": 8.122980362503737,
    "decode_energy_net_psys_j_q25": 7.770445141364107,
    "decode_energy_net_psys_j_q75": 8.295239841768808,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 2048,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 431.237221,
    "ttft_ms_q25": 402.07436699999994,
    "ttft_ms_q75": 438.01073199999996,
    "prefill_tokens_per_s": 4712.023686842189,
    "prefill_tokens_per_s_q25": 4640.265424769299,
    "prefill_tokens_per_s_q75": 5080.518769545417,
    "handoff_plus_first_decode_ms": 135.361615,
    "handoff_plus_first_decode_ms_q25": 135.25535100000002,
    "handoff_plus_first_decode_ms_q75": 192.700142,
    "steady_tpot_ms": 122.9227825,
    "steady_tpot_ms_q25": 122.83192475000001,
    "steady_tpot_ms_q75": 122.94218475,
    "steady_tpot_p95_ms": 123.09534180000001,
    "steady_tpot_p95_ms_q25": 123.087232275,
    "steady_tpot_p95_ms_q75": 124.40221297500001,
    "decode_tokens_per_s": 8.08838783595787,
    "decode_tokens_per_s_q25": 7.833662975686522,
    "decode_tokens_per_s_q75": 8.092821854246818,
    "decode_ms": 1854.510479,
    "decode_ms_q25": 1853.4949579999998,
    "decode_ms_q75": 1916.8397810000001,
    "end_to_end_ms": 2297.2636799999996,
    "end_to_end_ms_q25": 2291.50569,
    "end_to_end_ms_q75": 2324.672138,
    "output_tokens_per_s_e2e": 6.964807801253361,
    "output_tokens_per_s_e2e_q25": 6.883647894393295,
    "output_tokens_per_s_e2e_q75": 6.982352735749647,
    "all_tokens_per_s_e2e": 891.4953985604302,
    "all_tokens_per_s_e2e_q25": 881.1069304823418,
    "all_tokens_per_s_e2e_q75": 893.7411501759548,
    "energy_net_psys_j": 44.82079071975753,
    "energy_net_psys_j_q25": 44.10535636020846,
    "energy_net_psys_j_q75": 44.86099473089841,
    "prefill_energy_net_psys_j": 11.128946290697668,
    "prefill_energy_net_psys_j_q25": 10.176572668790886,
    "prefill_energy_net_psys_j_q75": 11.236914521531057,
    "decode_energy_net_psys_j": 33.55631598967485,
    "decode_energy_net_psys_j_q25": 32.908645849818285,
    "decode_energy_net_psys_j_q75": 34.57645383127414,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.27458772473171217,
    "energy_reduction_vs_gpu": -1.3888123426976549
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 2048,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 579.501888,
    "ttft_ms_q25": 575.784952,
    "ttft_ms_q75": 579.6454225,
    "prefill_tokens_per_s": 3506.459671793166,
    "prefill_tokens_per_s_q25": 3505.5916008689364,
    "prefill_tokens_per_s_q75": 3529.2424280775795,
    "handoff_plus_first_decode_ms": 135.361615,
    "handoff_plus_first_decode_ms_q25": 135.25535100000002,
    "handoff_plus_first_decode_ms_q75": 192.700142,
    "steady_tpot_ms": 122.9227825,
    "steady_tpot_ms_q25": 122.83192475000001,
    "steady_tpot_ms_q75": 122.94218475,
    "steady_tpot_p95_ms": 123.09534180000001,
    "steady_tpot_p95_ms_q25": 123.087232275,
    "steady_tpot_p95_ms_q75": 124.40221297500001,
    "decode_tokens_per_s": 8.08838783595787,
    "decode_tokens_per_s_q25": 7.833662975686522,
    "decode_tokens_per_s_q75": 8.092821854246818,
    "decode_ms": 1854.510479,
    "decode_ms_q25": 1853.4949579999998,
    "decode_ms_q75": 1916.8397810000001,
    "end_to_end_ms": 2431.9813249999997,
    "end_to_end_ms_q25": 2429.27991,
    "end_to_end_ms_q75": 2495.4696825,
    "output_tokens_per_s_e2e": 6.5789978876585335,
    "output_tokens_per_s_e2e_q25": 6.415771385561058,
    "output_tokens_per_s_e2e_q75": 6.58632202887849,
    "all_tokens_per_s_e2e": 842.1117296202923,
    "all_tokens_per_s_e2e_q25": 821.2187373518154,
    "all_tokens_per_s_e2e_q75": 843.0492196964467,
    "energy_net_psys_j": 43.18449896937483,
    "energy_net_psys_j_q25": 42.46109409245022,
    "energy_net_psys_j_q75": 44.329350493740165,
    "prefill_energy_net_psys_j": 9.628182979699977,
    "prefill_energy_net_psys_j_q25": 9.552448242631936,
    "prefill_energy_net_psys_j_q75": 9.752896662466021,
    "decode_energy_net_psys_j": 33.55631598967485,
    "decode_energy_net_psys_j_q25": 32.908645849818285,
    "decode_energy_net_psys_j_q75": 34.57645383127414,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.25937715907419645,
    "energy_reduction_vs_gpu": -1.3016029502081663
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 2048,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 579.501888,
    "ttft_ms_q25": 575.784952,
    "ttft_ms_q75": 579.6454225,
    "prefill_tokens_per_s": 3506.459671793166,
    "prefill_tokens_per_s_q25": 3505.5916008689364,
    "prefill_tokens_per_s_q75": 3529.2424280775795,
    "handoff_plus_first_decode_ms": 41.103348999999994,
    "handoff_plus_first_decode_ms_q25": 32.619724,
    "handoff_plus_first_decode_ms_q75": 46.111813,
    "steady_tpot_ms": 10.842264,
    "steady_tpot_ms_q25": 10.804337,
    "steady_tpot_ms_q75": 10.941298,
    "steady_tpot_p95_ms": 13.5699021,
    "steady_tpot_p95_ms_q25": 13.259286475,
    "steady_tpot_p95_ms_q75": 18.472543424999998,
    "decode_tokens_per_s": 70.3662392466489,
    "decode_tokens_per_s_q25": 67.74458303239297,
    "decode_tokens_per_s_q75": 75.50219641065138,
    "decode_ms": 213.170409,
    "decode_ms_q25": 199.59328649999998,
    "decode_ms_q75": 221.75201850000005,
    "end_to_end_ms": 792.959366,
    "end_to_end_ms_q25": 779.238709,
    "end_to_end_ms_q75": 797.680505,
    "output_tokens_per_s_e2e": 20.177578683142762,
    "output_tokens_per_s_e2e_q25": 20.058858645168172,
    "output_tokens_per_s_e2e_q75": 20.53922872585386,
    "all_tokens_per_s_e2e": 2582.7300714422736,
    "all_tokens_per_s_e2e_q25": 2567.533906581526,
    "all_tokens_per_s_e2e_q75": 2629.021276909294,
    "energy_net_psys_j": 17.599693868067632,
    "energy_net_psys_j_q25": 17.322893383996043,
    "energy_net_psys_j_q75": 17.97240176716679,
    "prefill_energy_net_psys_j": 9.628182979699977,
    "prefill_energy_net_psys_j_q25": 9.552448242631936,
    "prefill_energy_net_psys_j_q75": 9.752896662466021,
    "decode_energy_net_psys_j": 8.122980362503737,
    "decode_energy_net_psys_j_q25": 7.770445141364107,
    "decode_energy_net_psys_j_q75": 8.295239841768808,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.7955015528500611,
    "energy_reduction_vs_gpu": 0.0619896422039824
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 4096,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1081.158091,
    "ttft_ms_q25": 1026.184426,
    "ttft_ms_q75": 1081.21848,
    "prefill_tokens_per_s": 3773.731181372623,
    "prefill_tokens_per_s_q25": 3773.5204199688396,
    "prefill_tokens_per_s_q75": 3987.336551180539,
    "handoff_plus_first_decode_ms": 35.639419,
    "handoff_plus_first_decode_ms_q25": 35.1314085,
    "handoff_plus_first_decode_ms_q75": 52.99225949999999,
    "steady_tpot_ms": 13.736344500000001,
    "steady_tpot_ms_q25": 13.726874500000001,
    "steady_tpot_ms_q75": 13.853233750000001,
    "steady_tpot_p95_ms": 17.34561505,
    "steady_tpot_p95_ms_q25": 17.285200749999998,
    "steady_tpot_p95_ms_q75": 17.7116704,
    "decode_tokens_per_s": 61.99981878692965,
    "decode_tokens_per_s_q25": 58.62843857291044,
    "decode_tokens_per_s_q75": 62.495546624094764,
    "decode_ms": 241.936191,
    "decode_ms_q25": 240.0322050000001,
    "decode_ms_q75": 256.697362,
    "end_to_end_ms": 1319.2863100000002,
    "end_to_end_ms_q25": 1280.977802,
    "end_to_end_ms_q75": 1321.250685,
    "output_tokens_per_s_e2e": 12.12776929368728,
    "output_tokens_per_s_e2e_q25": 12.109765049686093,
    "output_tokens_per_s_e2e_q75": 12.501639276033021,
    "all_tokens_per_s_e2e": 3104.7089391839436,
    "all_tokens_per_s_e2e_q25": 3100.09985271964,
    "all_tokens_per_s_e2e_q75": 3200.4196546644534,
    "energy_net_psys_j": 35.23845258394553,
    "energy_net_psys_j_q25": 34.750285630277006,
    "energy_net_psys_j_q75": 35.414115039388406,
    "prefill_energy_net_psys_j": 27.349379707607135,
    "prefill_energy_net_psys_j_q25": 25.705690942943104,
    "prefill_energy_net_psys_j_q75": 27.364781005554192,
    "decode_energy_net_psys_j": 8.240397787224154,
    "decode_energy_net_psys_j_q25": 8.049334033834215,
    "decode_energy_net_psys_j_q75": 9.220257142776784,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 4096,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1081.158091,
    "ttft_ms_q25": 1026.184426,
    "ttft_ms_q75": 1081.21848,
    "prefill_tokens_per_s": 3773.731181372623,
    "prefill_tokens_per_s_q25": 3773.5204199688396,
    "prefill_tokens_per_s_q75": 3987.336551180539,
    "handoff_plus_first_decode_ms": 173.10389700000002,
    "handoff_plus_first_decode_ms_q25": 162.950516,
    "handoff_plus_first_decode_ms_q75": 224.01592950000003,
    "steady_tpot_ms": 125.98738949999999,
    "steady_tpot_ms_q25": 125.96643800000001,
    "steady_tpot_ms_q75": 126.00246949999999,
    "steady_tpot_p95_ms": 134.6569466,
    "steady_tpot_p95_ms_q25": 130.53024165,
    "steady_tpot_p95_ms_q75": 135.23860355,
    "decode_tokens_per_s": 7.639847760875907,
    "decode_tokens_per_s_q25": 7.456068285060912,
    "decode_tokens_per_s_q75": 7.739068646464693,
    "decode_ms": 1963.389909,
    "decode_ms_q25": 1938.5363634999999,
    "decode_ms_q75": 2013.007131,
    "end_to_end_ms": 3033.8351140000004,
    "end_to_end_ms_q25": 3014.3984005,
    "end_to_end_ms_q75": 3039.191557,
    "output_tokens_per_s_e2e": 5.273852862393891,
    "output_tokens_per_s_e2e_q25": 5.264574279087667,
    "output_tokens_per_s_e2e_q75": 5.308079132323302,
    "all_tokens_per_s_e2e": 1350.1063327728361,
    "all_tokens_per_s_e2e_q25": 1347.7310154464428,
    "all_tokens_per_s_e2e_q75": 1358.8682578747653,
    "energy_net_psys_j": 62.1721427445659,
    "energy_net_psys_j_q25": 60.38326964061885,
    "energy_net_psys_j_q75": 68.77355094114523,
    "prefill_energy_net_psys_j": 27.349379707607135,
    "prefill_energy_net_psys_j_q25": 25.705690942943104,
    "prefill_energy_net_psys_j_q75": 27.364781005554192,
    "decode_energy_net_psys_j": 38.11014056628683,
    "decode_energy_net_psys_j_q25": 34.67757869767575,
    "decode_energy_net_psys_j_q75": 43.05245870025506,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.434857617644411,
    "energy_reduction_vs_gpu": -0.7643266995467115
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 4096,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1298.635947,
    "ttft_ms_q25": 1293.3136195000002,
    "ttft_ms_q75": 1306.073258,
    "prefill_tokens_per_s": 3141.7580958122053,
    "prefill_tokens_per_s_q25": 3123.9689492329644,
    "prefill_tokens_per_s_q75": 3154.740687865078,
    "handoff_plus_first_decode_ms": 173.10389700000002,
    "handoff_plus_first_decode_ms_q25": 162.950516,
    "handoff_plus_first_decode_ms_q75": 224.01592950000003,
    "steady_tpot_ms": 125.98738949999999,
    "steady_tpot_ms_q25": 125.96643800000001,
    "steady_tpot_ms_q75": 126.00246949999999,
    "steady_tpot_p95_ms": 134.6569466,
    "steady_tpot_p95_ms_q25": 130.53024165,
    "steady_tpot_p95_ms_q75": 135.23860355,
    "decode_tokens_per_s": 7.639847760875907,
    "decode_tokens_per_s_q25": 7.456068285060912,
    "decode_tokens_per_s_q75": 7.739068646464693,
    "decode_ms": 1963.389909,
    "decode_ms_q25": 1938.5363634999999,
    "decode_ms_q75": 2013.007131,
    "end_to_end_ms": 3276.900478,
    "end_to_end_ms_q25": 3244.6096215,
    "end_to_end_ms_q75": 3313.7580615,
    "output_tokens_per_s_e2e": 4.882662780703467,
    "output_tokens_per_s_e2e_q25": 4.828952307701088,
    "output_tokens_per_s_e2e_q75": 4.931744261939837,
    "all_tokens_per_s_e2e": 1249.9616718600876,
    "all_tokens_per_s_e2e_q25": 1236.2117907714785,
    "all_tokens_per_s_e2e_q75": 1262.5265310565983,
    "energy_net_psys_j": 64.77302497724148,
    "energy_net_psys_j_q25": 59.503293836218525,
    "energy_net_psys_j_q75": 73.93139059507189,
    "prefill_energy_net_psys_j": 26.662884410954646,
    "prefill_energy_net_psys_j_q25": 24.825715138542776,
    "prefill_energy_net_psys_j_q75": 30.87893189481683,
    "decode_energy_net_psys_j": 38.11014056628683,
    "decode_energy_net_psys_j_q25": 34.67757869767575,
    "decode_energy_net_psys_j_q75": 43.05245870025506,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.4026018851830386,
    "energy_reduction_vs_gpu": -0.8381347711832203
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 4096,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1298.635947,
    "ttft_ms_q25": 1293.3136195000002,
    "ttft_ms_q75": 1306.073258,
    "prefill_tokens_per_s": 3141.7580958122053,
    "prefill_tokens_per_s_q25": 3123.9689492329644,
    "prefill_tokens_per_s_q75": 3154.740687865078,
    "handoff_plus_first_decode_ms": 35.639419,
    "handoff_plus_first_decode_ms_q25": 35.1314085,
    "handoff_plus_first_decode_ms_q75": 52.99225949999999,
    "steady_tpot_ms": 13.736344500000001,
    "steady_tpot_ms_q25": 13.726874500000001,
    "steady_tpot_ms_q75": 13.853233750000001,
    "steady_tpot_p95_ms": 17.34561505,
    "steady_tpot_p95_ms_q25": 17.285200749999998,
    "steady_tpot_p95_ms_q75": 17.7116704,
    "decode_tokens_per_s": 61.99981878692965,
    "decode_tokens_per_s_q25": 58.62843857291044,
    "decode_tokens_per_s_q75": 62.495546624094764,
    "decode_ms": 241.936191,
    "decode_ms_q25": 240.0322050000001,
    "decode_ms_q75": 256.697362,
    "end_to_end_ms": 1551.6387880000002,
    "end_to_end_ms_q25": 1546.105463,
    "end_to_end_ms_q75": 1555.5443065000002,
    "output_tokens_per_s_e2e": 10.311678287330878,
    "output_tokens_per_s_e2e_q25": 10.285853506262837,
    "output_tokens_per_s_e2e_q75": 10.348715090640408,
    "all_tokens_per_s_e2e": 2639.7896415567047,
    "all_tokens_per_s_e2e_q25": 2633.1784976032864,
    "all_tokens_per_s_e2e_q75": 2649.2710632039443,
    "energy_net_psys_j": 36.86300090928406,
    "energy_net_psys_j_q25": 34.045972281319564,
    "energy_net_psys_j_q75": 39.90812528420368,
    "prefill_energy_net_psys_j": 26.662884410954646,
    "prefill_energy_net_psys_j_q25": 24.825715138542776,
    "prefill_energy_net_psys_j_q75": 30.87893189481683,
    "decode_energy_net_psys_j": 8.240397787224154,
    "decode_energy_net_psys_j_q25": 8.049334033834215,
    "decode_energy_net_psys_j_q75": 9.220257142776784,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.8502534998499922,
    "energy_reduction_vs_gpu": -0.04610157955910554
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 8192,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2767.602321,
    "ttft_ms_q25": 2663.1035624999995,
    "ttft_ms_q75": 2778.9175015,
    "prefill_tokens_per_s": 2954.1816531812337,
    "prefill_tokens_per_s_q25": 2942.2016133619118,
    "prefill_tokens_per_s_q75": 3074.836595538216,
    "handoff_plus_first_decode_ms": 49.816934,
    "handoff_plus_first_decode_ms_q25": 49.792398500000004,
    "handoff_plus_first_decode_ms_q75": 56.983838,
    "steady_tpot_ms": 16.926910499999998,
    "steady_tpot_ms_q25": 16.8871495,
    "steady_tpot_ms_q75": 16.975437,
    "steady_tpot_p95_ms": 20.88518585,
    "steady_tpot_p95_ms_q25": 20.627281875,
    "steady_tpot_p95_ms_q75": 21.275896875,
    "decode_tokens_per_s": 49.70044004073806,
    "decode_tokens_per_s_q25": 49.08083068974983,
    "decode_tokens_per_s_q75": 50.05398007053218,
    "decode_ms": 301.8081929999994,
    "decode_ms_q25": 299.6914199999999,
    "decode_ms_q75": 305.6670139999999,
    "end_to_end_ms": 3065.176968,
    "end_to_end_ms_q25": 2966.6538035000003,
    "end_to_end_ms_q75": 3078.6089215,
    "output_tokens_per_s_e2e": 5.219926995092832,
    "output_tokens_per_s_e2e_q25": 5.197251416241864,
    "output_tokens_per_s_e2e_q75": 5.399236723534921,
    "all_tokens_per_s_e2e": 2672.60262148753,
    "all_tokens_per_s_e2e_q25": 2660.992725115834,
    "all_tokens_per_s_e2e_q75": 2764.4092024498796,
    "energy_net_psys_j": 83.56347810145249,
    "energy_net_psys_j_q25": 80.89226014035711,
    "energy_net_psys_j_q75": 83.82630873998627,
    "prefill_energy_net_psys_j": 73.88001436422923,
    "prefill_energy_net_psys_j_q25": 70.81250135250667,
    "prefill_energy_net_psys_j_q75": 74.33757825423756,
    "decode_energy_net_psys_j": 9.683463737223263,
    "decode_energy_net_psys_j_q25": 9.488730485748711,
    "decode_energy_net_psys_j_q75": 10.079758787850423,
    "fastest_e2e": true,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 8192,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2767.602321,
    "ttft_ms_q25": 2663.1035624999995,
    "ttft_ms_q75": 2778.9175015,
    "prefill_tokens_per_s": 2954.1816531812337,
    "prefill_tokens_per_s_q25": 2942.2016133619118,
    "prefill_tokens_per_s_q75": 3074.836595538216,
    "handoff_plus_first_decode_ms": 193.19836700000002,
    "handoff_plus_first_decode_ms_q25": 191.31255650000003,
    "handoff_plus_first_decode_ms_q75": 239.21696500000002,
    "steady_tpot_ms": 128.32535049999998,
    "steady_tpot_ms_q25": 128.0788365,
    "steady_tpot_ms_q75": 128.34406925,
    "steady_tpot_p95_ms": 128.76579915,
    "steady_tpot_p95_ms_q25": 128.42559725,
    "steady_tpot_p95_ms_q75": 133.27836625,
    "decode_tokens_per_s": 7.538911147524052,
    "decode_tokens_per_s_q25": 7.328415213156733,
    "decode_tokens_per_s_q75": 7.561856309187208,
    "decode_ms": 1989.6772499999997,
    "decode_ms_q25": 1983.6581779999997,
    "decode_ms_q75": 2048.5173345,
    "end_to_end_ms": 4745.241426999999,
    "end_to_end_ms_q25": 4705.601825,
    "end_to_end_ms_q75": 4762.5756795,
    "output_tokens_per_s_e2e": 3.371798937133405,
    "output_tokens_per_s_e2e_q25": 3.359571172387418,
    "output_tokens_per_s_e2e_q75": 3.400443997145498,
    "all_tokens_per_s_e2e": 1726.3610558123034,
    "all_tokens_per_s_e2e_q25": 1720.100440262358,
    "all_tokens_per_s_e2e_q75": 1741.027326538495,
    "energy_net_psys_j": 104.69055899941593,
    "energy_net_psys_j_q25": 103.3705781468208,
    "energy_net_psys_j_q75": 105.1172365495738,
    "prefill_energy_net_psys_j": 73.88001436422923,
    "prefill_energy_net_psys_j_q25": 70.81250135250667,
    "prefill_energy_net_psys_j_q75": 74.33757825423756,
    "decode_energy_net_psys_j": 30.8105446351867,
    "decode_energy_net_psys_j_q25": 30.779658295336247,
    "decode_energy_net_psys_j_q75": 32.55807679431413,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.6459475276767621,
    "energy_reduction_vs_gpu": -0.2528267297863491
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 8192,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3803.3502559999997,
    "ttft_ms_q25": 3802.8790125,
    "ttft_ms_q75": 3813.0174985,
    "prefill_tokens_per_s": 2149.683686665959,
    "prefill_tokens_per_s_q25": 2144.247320628655,
    "prefill_tokens_per_s_q75": 2149.9501032435437,
    "handoff_plus_first_decode_ms": 193.19836700000002,
    "handoff_plus_first_decode_ms_q25": 191.31255650000003,
    "handoff_plus_first_decode_ms_q75": 239.21696500000002,
    "steady_tpot_ms": 128.32535049999998,
    "steady_tpot_ms_q25": 128.0788365,
    "steady_tpot_ms_q75": 128.34406925,
    "steady_tpot_p95_ms": 128.76579915,
    "steady_tpot_p95_ms_q25": 128.42559725,
    "steady_tpot_p95_ms_q75": 133.27836625,
    "decode_tokens_per_s": 7.538911147524052,
    "decode_tokens_per_s_q25": 7.328415213156733,
    "decode_tokens_per_s_q75": 7.561856309187208,
    "decode_ms": 1989.6772499999997,
    "decode_ms_q25": 1983.6581779999997,
    "decode_ms_q75": 2048.5173345,
    "end_to_end_ms": 5793.027505999999,
    "end_to_end_ms_q25": 5786.537190499999,
    "end_to_end_ms_q75": 5861.534833,
    "output_tokens_per_s_e2e": 2.7619409684190788,
    "output_tokens_per_s_e2e_q25": 2.7300334055429687,
    "output_tokens_per_s_e2e_q75": 2.7650423045613426,
    "all_tokens_per_s_e2e": 1414.1137758305683,
    "all_tokens_per_s_e2e_q25": 1397.777103638,
    "all_tokens_per_s_e2e_q75": 1415.7016599354074,
    "energy_net_psys_j": 92.38107921915228,
    "energy_net_psys_j_q25": 92.31603133301135,
    "energy_net_psys_j_q75": 96.454425338855,
    "prefill_energy_net_psys_j": 61.63230726366649,
    "prefill_energy_net_psys_j_q25": 61.536373037675105,
    "prefill_energy_net_psys_j_q75": 63.92723488439133,
    "decode_energy_net_psys_j": 30.8105446351867,
    "decode_energy_net_psys_j_q25": 30.779658295336247,
    "decode_energy_net_psys_j_q75": 32.55807679431413,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.529114865210861,
    "energy_reduction_vs_gpu": -0.10551979546608337
  },
  {
    "precision": "INT8",
    "model": "qwen25_0_5b",
    "model_label": "Qwen2.5 0.5B",
    "family": "Qwen2.5",
    "size_b": 0.5,
    "context_length": 8192,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3803.3502559999997,
    "ttft_ms_q25": 3802.8790125,
    "ttft_ms_q75": 3813.0174985,
    "prefill_tokens_per_s": 2149.683686665959,
    "prefill_tokens_per_s_q25": 2144.247320628655,
    "prefill_tokens_per_s_q75": 2149.9501032435437,
    "handoff_plus_first_decode_ms": 49.816934,
    "handoff_plus_first_decode_ms_q25": 49.792398500000004,
    "handoff_plus_first_decode_ms_q75": 56.983838,
    "steady_tpot_ms": 16.926910499999998,
    "steady_tpot_ms_q25": 16.8871495,
    "steady_tpot_ms_q75": 16.975437,
    "steady_tpot_p95_ms": 20.88518585,
    "steady_tpot_p95_ms_q25": 20.627281875,
    "steady_tpot_p95_ms_q75": 21.275896875,
    "decode_tokens_per_s": 49.70044004073806,
    "decode_tokens_per_s_q25": 49.08083068974983,
    "decode_tokens_per_s_q75": 50.05398007053218,
    "decode_ms": 301.8081929999994,
    "decode_ms_q25": 299.6914199999999,
    "decode_ms_q75": 305.6670139999999,
    "end_to_end_ms": 4105.158448999999,
    "end_to_end_ms_q25": 4102.5704325,
    "end_to_end_ms_q75": 4118.6845125,
    "output_tokens_per_s_e2e": 3.8975353080214337,
    "output_tokens_per_s_e2e_q25": 3.8847774126769945,
    "output_tokens_per_s_e2e_q75": 3.8999955345939004,
    "all_tokens_per_s_e2e": 1995.538077706974,
    "all_tokens_per_s_e2e_q25": 1989.0060352906212,
    "all_tokens_per_s_e2e_q75": 1996.797713712077,
    "energy_net_psys_j": 71.12390254890698,
    "energy_net_psys_j_q25": 71.02510352342381,
    "energy_net_psys_j_q75": 73.91105944625036,
    "prefill_energy_net_psys_j": 61.63230726366649,
    "prefill_energy_net_psys_j_q25": 61.536373037675105,
    "prefill_energy_net_psys_j_q75": 63.92723488439133,
    "decode_energy_net_psys_j": 9.683463737223263,
    "decode_energy_net_psys_j_q25": 9.488730485748711,
    "decode_energy_net_psys_j_q75": 10.079758787850423,
    "fastest_e2e": false,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 0.7466647161321303,
    "energy_reduction_vs_gpu": 0.1488637839780066
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 1024,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 339.972811,
    "ttft_ms_q25": 317.2443015,
    "ttft_ms_q75": 342.7688715,
    "prefill_tokens_per_s": 2964.9429818668646,
    "prefill_tokens_per_s_q25": 2940.9528162317993,
    "prefill_tokens_per_s_q75": 3193.754938851227,
    "handoff_plus_first_decode_ms": 35.874994,
    "handoff_plus_first_decode_ms_q25": 33.997557,
    "handoff_plus_first_decode_ms_q75": 41.853654500000005,
    "steady_tpot_ms": 23.400827999999997,
    "steady_tpot_ms_q25": 23.312720249999998,
    "steady_tpot_ms_q75": 23.736918,
    "steady_tpot_p95_ms": 24.7919254,
    "steady_tpot_p95_ms_q25": 24.539722849999997,
    "steady_tpot_p95_ms_q75": 30.054694724999997,
    "decode_tokens_per_s": 39.66048645855221,
    "decode_tokens_per_s_q25": 38.80590180853985,
    "decode_tokens_per_s_q75": 40.35164683844978,
    "decode_ms": 378.21018699999996,
    "decode_ms_q25": 371.841132,
    "decode_ms_q75": 386.72669300000007,
    "end_to_end_ms": 705.444888,
    "end_to_end_ms_q25": 689.0854334999999,
    "end_to_end_ms_q75": 723.1265095000001,
    "output_tokens_per_s_e2e": 22.680722863215365,
    "output_tokens_per_s_e2e_q25": 22.139378971008853,
    "output_tokens_per_s_e2e_q75": 23.232276192937274,
    "all_tokens_per_s_e2e": 1451.5662632457834,
    "all_tokens_per_s_e2e_q25": 1416.9202541445666,
    "all_tokens_per_s_e2e_q75": 1486.8656763479855,
    "energy_net_psys_j": 23.167475685392333,
    "energy_net_psys_j_q25": 23.12918756901884,
    "energy_net_psys_j_q75": 23.169200101865627,
    "prefill_energy_net_psys_j": 9.813751017035404,
    "prefill_energy_net_psys_j_q25": 9.419546391071115,
    "prefill_energy_net_psys_j_q75": 9.838721782282576,
    "decode_energy_net_psys_j": 13.303783137862586,
    "decode_energy_net_psys_j_q25": 13.290465786736267,
    "decode_energy_net_psys_j_q75": 13.72468294554734,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 1024,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 339.972811,
    "ttft_ms_q25": 317.2443015,
    "ttft_ms_q75": 342.7688715,
    "prefill_tokens_per_s": 2964.9429818668646,
    "prefill_tokens_per_s_q25": 2940.9528162317993,
    "prefill_tokens_per_s_q75": 3193.754938851227,
    "handoff_plus_first_decode_ms": 508.60053000000005,
    "handoff_plus_first_decode_ms_q25": 505.48739,
    "handoff_plus_first_decode_ms_q75": 592.453497,
    "steady_tpot_ms": 504.70483099999996,
    "steady_tpot_ms_q25": 504.49935800000003,
    "steady_tpot_ms_q75": 504.81454199999996,
    "steady_tpot_p95_ms": 505.11252255000005,
    "steady_tpot_p95_ms_q25": 505.0448423750001,
    "steady_tpot_p95_ms_q75": 505.95412972500003,
    "decode_tokens_per_s": 1.9808261869857582,
    "decode_tokens_per_s_q25": 1.9584322120707864,
    "decode_tokens_per_s_q75": 1.9849727999164217,
    "decode_ms": 7572.597787,
    "decode_ms_q25": 7556.811589499999,
    "decode_ms_q75": 7660.189324000001,
    "end_to_end_ms": 7912.570597999999,
    "end_to_end_ms_q25": 7899.580461,
    "end_to_end_ms_q75": 7977.4336255,
    "output_tokens_per_s_e2e": 2.0220988617838302,
    "output_tokens_per_s_e2e_q25": 2.005790155485278,
    "output_tokens_per_s_e2e_q75": 2.025429495196115,
    "all_tokens_per_s_e2e": 129.41432715416514,
    "all_tokens_per_s_e2e_q25": 128.3705699510578,
    "all_tokens_per_s_e2e_q75": 129.62748769255137,
    "energy_net_psys_j": 107.2488111607607,
    "energy_net_psys_j_q25": 106.24865083453247,
    "energy_net_psys_j_q75": 109.9971724144817,
    "prefill_energy_net_psys_j": 9.813751017035404,
    "prefill_energy_net_psys_j_q25": 9.419546391071115,
    "prefill_energy_net_psys_j_q75": 9.838721782282576,
    "decode_energy_net_psys_j": 97.38511861323094,
    "decode_energy_net_psys_j_q25": 96.4099290522499,
    "decode_energy_net_psys_j_q75": 100.55265525816341,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.08915495656725135,
    "energy_reduction_vs_gpu": -3.62928342376048
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 1024,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 910.8172890000001,
    "ttft_ms_q25": 909.878027,
    "ttft_ms_q75": 912.985671,
    "prefill_tokens_per_s": 1106.698360004451,
    "prefill_tokens_per_s_q25": 1104.076129577348,
    "prefill_tokens_per_s_q75": 1107.8419790853409,
    "handoff_plus_first_decode_ms": 508.60053000000005,
    "handoff_plus_first_decode_ms_q25": 505.48739,
    "handoff_plus_first_decode_ms_q75": 592.453497,
    "steady_tpot_ms": 504.70483099999996,
    "steady_tpot_ms_q25": 504.49935800000003,
    "steady_tpot_ms_q75": 504.81454199999996,
    "steady_tpot_p95_ms": 505.11252255000005,
    "steady_tpot_p95_ms_q25": 505.0448423750001,
    "steady_tpot_p95_ms_q75": 505.95412972500003,
    "decode_tokens_per_s": 1.9808261869857582,
    "decode_tokens_per_s_q25": 1.9584322120707864,
    "decode_tokens_per_s_q75": 1.9849727999164217,
    "decode_ms": 7572.597787,
    "decode_ms_q25": 7556.811589499999,
    "decode_ms_q75": 7660.189324000001,
    "end_to_end_ms": 8481.536552,
    "end_to_end_ms_q25": 8468.8579985,
    "end_to_end_ms_q75": 8570.067351000002,
    "output_tokens_per_s_e2e": 1.8864506333144433,
    "output_tokens_per_s_e2e_q25": 1.8671624091760606,
    "output_tokens_per_s_e2e_q75": 1.889279034182134,
    "all_tokens_per_s_e2e": 120.73284053212437,
    "all_tokens_per_s_e2e_q25": 119.49839418726788,
    "all_tokens_per_s_e2e_q75": 120.91385818765657,
    "energy_net_psys_j": 112.08481262627237,
    "energy_net_psys_j_q25": 110.56533729627971,
    "energy_net_psys_j_q75": 115.20917096732981,
    "prefill_energy_net_psys_j": 14.61333740529138,
    "prefill_energy_net_psys_j_q25": 14.112229940154794,
    "prefill_energy_net_psys_j_q75": 14.6565157091664,
    "decode_energy_net_psys_j": 97.38511861323094,
    "decode_energy_net_psys_j_q25": 96.4099290522499,
    "decode_energy_net_psys_j_q75": 100.55265525816341,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.08317418473350228,
    "energy_reduction_vs_gpu": -3.8380243988750413
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 1024,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 910.8172890000001,
    "ttft_ms_q25": 909.878027,
    "ttft_ms_q75": 912.985671,
    "prefill_tokens_per_s": 1106.698360004451,
    "prefill_tokens_per_s_q25": 1104.076129577348,
    "prefill_tokens_per_s_q75": 1107.8419790853409,
    "handoff_plus_first_decode_ms": 35.874994,
    "handoff_plus_first_decode_ms_q25": 33.997557,
    "handoff_plus_first_decode_ms_q75": 41.853654500000005,
    "steady_tpot_ms": 23.400827999999997,
    "steady_tpot_ms_q25": 23.312720249999998,
    "steady_tpot_ms_q75": 23.736918,
    "steady_tpot_p95_ms": 24.7919254,
    "steady_tpot_p95_ms_q25": 24.539722849999997,
    "steady_tpot_p95_ms_q75": 30.054694724999997,
    "decode_tokens_per_s": 39.66048645855221,
    "decode_tokens_per_s_q25": 38.80590180853985,
    "decode_tokens_per_s_q75": 40.35164683844978,
    "decode_ms": 378.21018699999996,
    "decode_ms_q25": 371.841132,
    "decode_ms_q75": 386.72669300000007,
    "end_to_end_ms": 1289.027476,
    "end_to_end_ms_q25": 1281.719159,
    "end_to_end_ms_q75": 1299.712364,
    "output_tokens_per_s_e2e": 12.412458460272573,
    "output_tokens_per_s_e2e_q25": 12.311248137791932,
    "output_tokens_per_s_e2e_q75": 12.48363972944213,
    "all_tokens_per_s_e2e": 794.3973414574447,
    "all_tokens_per_s_e2e_q25": 787.9198808186836,
    "all_tokens_per_s_e2e_q75": 798.9529426842963,
    "energy_net_psys_j": 28.003477150904004,
    "energy_net_psys_j_q25": 27.445874030766078,
    "energy_net_psys_j_q75": 28.38119865471374,
    "prefill_energy_net_psys_j": 14.61333740529138,
    "prefill_energy_net_psys_j_q25": 14.112229940154794,
    "prefill_energy_net_psys_j_q75": 14.6565157091664,
    "decode_energy_net_psys_j": 13.303783137862586,
    "decode_energy_net_psys_j_q25": 13.290465786736267,
    "decode_energy_net_psys_j_q75": 13.72468294554734,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.5472690855194773,
    "energy_reduction_vs_gpu": -0.2087409751145608
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 2048,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 699.393099,
    "ttft_ms_q25": 672.2121945,
    "ttft_ms_q75": 711.764133,
    "prefill_tokens_per_s": 2905.376108093397,
    "prefill_tokens_per_s_q25": 2855.741028547521,
    "prefill_tokens_per_s_q75": 3027.805440987726,
    "handoff_plus_first_decode_ms": 42.768417,
    "handoff_plus_first_decode_ms_q25": 41.313729499999994,
    "handoff_plus_first_decode_ms_q75": 48.019420499999995,
    "steady_tpot_ms": 26.215950000000003,
    "steady_tpot_ms_q25": 26.068020500000003,
    "steady_tpot_ms_q75": 26.273407750000004,
    "steady_tpot_p95_ms": 29.713804399999997,
    "steady_tpot_p95_ms_q25": 29.385052249999998,
    "steady_tpot_p95_ms_q75": 30.03668225,
    "decode_tokens_per_s": 35.504392032945084,
    "decode_tokens_per_s_q25": 35.41831321350345,
    "decode_tokens_per_s_q75": 35.65444674854093,
    "decode_ms": 422.482942,
    "decode_ms_q25": 420.71233899999993,
    "decode_ms_q75": 423.51222400000006,
    "end_to_end_ms": 1118.3348349999999,
    "end_to_end_ms_q25": 1093.9538155,
    "end_to_end_ms_q75": 1132.4764719999998,
    "output_tokens_per_s_e2e": 14.306985259919943,
    "output_tokens_per_s_e2e_q25": 14.130532271324993,
    "output_tokens_per_s_e2e_q75": 14.63311442841865,
    "all_tokens_per_s_e2e": 1831.2941132697526,
    "all_tokens_per_s_e2e_q25": 1808.7081307295991,
    "all_tokens_per_s_e2e_q75": 1873.0386468375873,
    "energy_net_psys_j": 35.91370598660502,
    "energy_net_psys_j_q25": 35.260720895451456,
    "energy_net_psys_j_q75": 35.943466413730775,
    "prefill_energy_net_psys_j": 21.3310625709722,
    "prefill_energy_net_psys_j_q25": 20.44179675025776,
    "prefill_energy_net_psys_j_q75": 21.336014385710566,
    "decode_energy_net_psys_j": 14.632260640407594,
    "decode_energy_net_psys_j_q25": 14.607452028020205,
    "decode_energy_net_psys_j_q75": 14.843732757581087,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 2048,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 699.393099,
    "ttft_ms_q25": 672.2121945,
    "ttft_ms_q75": 711.764133,
    "prefill_tokens_per_s": 2905.376108093397,
    "prefill_tokens_per_s_q25": 2855.741028547521,
    "prefill_tokens_per_s_q75": 3027.805440987726,
    "handoff_plus_first_decode_ms": 535.328571,
    "handoff_plus_first_decode_ms_q25": 533.273764,
    "handoff_plus_first_decode_ms_q75": 609.1082054999999,
    "steady_tpot_ms": 506.14783600000004,
    "steady_tpot_ms_q25": 505.94752200000005,
    "steady_tpot_ms_q75": 506.19152025000005,
    "steady_tpot_p95_ms": 506.6745119,
    "steady_tpot_p95_ms_q25": 506.631695675,
    "steady_tpot_p95_ms_q75": 507.505786525,
    "decode_tokens_per_s": 1.9677838729826396,
    "decode_tokens_per_s_q25": 1.953004612704816,
    "decode_tokens_per_s_q75": 1.968569280826768,
    "decode_ms": 7622.788359000001,
    "decode_ms_q25": 7619.748278000001,
    "decode_ms_q75": 7680.913268500002,
    "end_to_end_ms": 8340.843364,
    "end_to_end_ms_q25": 8331.512411,
    "end_to_end_ms_q75": 8362.456416,
    "output_tokens_per_s_e2e": 1.9182712468930618,
    "output_tokens_per_s_e2e_q25": 1.9133261905135255,
    "output_tokens_per_s_e2e_q75": 1.9204220410011144,
    "all_tokens_per_s_e2e": 245.53871960231191,
    "all_tokens_per_s_e2e_q25": 244.90575238573126,
    "all_tokens_per_s_e2e_q75": 245.81402124814264,
    "energy_net_psys_j": 120.36931267849182,
    "energy_net_psys_j_q25": 120.20880599990309,
    "energy_net_psys_j_q75": 120.92286109376958,
    "prefill_energy_net_psys_j": 21.3310625709722,
    "prefill_energy_net_psys_j_q25": 20.44179675025776,
    "prefill_energy_net_psys_j_q75": 21.336014385710566,
    "decode_energy_net_psys_j": 99.02834647804289,
    "decode_energy_net_psys_j_q25": 98.87279161419252,
    "decode_energy_net_psys_j_q75": 100.47611252877346,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.13407934739871227,
    "energy_reduction_vs_gpu": -2.351626053946836
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 2048,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1825.8328640000002,
    "ttft_ms_q25": 1822.8377095,
    "ttft_ms_q75": 1826.8214560000001,
    "prefill_tokens_per_s": 1112.9167625717575,
    "prefill_tokens_per_s_q25": 1112.3148287934919,
    "prefill_tokens_per_s_q75": 1114.7484362186804,
    "handoff_plus_first_decode_ms": 535.328571,
    "handoff_plus_first_decode_ms_q25": 533.273764,
    "handoff_plus_first_decode_ms_q75": 609.1082054999999,
    "steady_tpot_ms": 506.14783600000004,
    "steady_tpot_ms_q25": 505.94752200000005,
    "steady_tpot_ms_q75": 506.19152025000005,
    "steady_tpot_p95_ms": 506.6745119,
    "steady_tpot_p95_ms_q25": 506.631695675,
    "steady_tpot_p95_ms_q75": 507.505786525,
    "decode_tokens_per_s": 1.9677838729826396,
    "decode_tokens_per_s_q25": 1.953004612704816,
    "decode_tokens_per_s_q75": 1.968569280826768,
    "decode_ms": 7622.788359000001,
    "decode_ms_q25": 7619.748278000001,
    "decode_ms_q75": 7680.913268500002,
    "end_to_end_ms": 9442.630914000001,
    "end_to_end_ms_q25": 9442.5859875,
    "end_to_end_ms_q75": 9504.739570000002,
    "output_tokens_per_s_e2e": 1.6944430154818184,
    "output_tokens_per_s_e2e_q25": 1.6834425714615868,
    "output_tokens_per_s_e2e_q75": 1.694451077442433,
    "all_tokens_per_s_e2e": 216.88870598167276,
    "all_tokens_per_s_e2e_q25": 215.4806491470831,
    "all_tokens_per_s_e2e_q75": 216.88973791263143,
    "energy_net_psys_j": 130.28991509853228,
    "energy_net_psys_j_q25": 130.01844945518496,
    "energy_net_psys_j_q75": 132.09968109580248,
    "prefill_energy_net_psys_j": 31.261568620489392,
    "prefill_energy_net_psys_j_q25": 31.145657840992428,
    "prefill_energy_net_psys_j_q75": 31.623568567029025,
    "decode_energy_net_psys_j": 99.02834647804289,
    "decode_energy_net_psys_j_q25": 98.87279161419252,
    "decode_energy_net_psys_j_q75": 100.47611252877346,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.11843466563348509,
    "energy_reduction_vs_gpu": -2.6278604927914544
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 2048,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1825.8328640000002,
    "ttft_ms_q25": 1822.8377095,
    "ttft_ms_q75": 1826.8214560000001,
    "prefill_tokens_per_s": 1112.9167625717575,
    "prefill_tokens_per_s_q25": 1112.3148287934919,
    "prefill_tokens_per_s_q75": 1114.7484362186804,
    "handoff_plus_first_decode_ms": 42.768417,
    "handoff_plus_first_decode_ms_q25": 41.313729499999994,
    "handoff_plus_first_decode_ms_q75": 48.019420499999995,
    "steady_tpot_ms": 26.215950000000003,
    "steady_tpot_ms_q25": 26.068020500000003,
    "steady_tpot_ms_q75": 26.273407750000004,
    "steady_tpot_p95_ms": 29.713804399999997,
    "steady_tpot_p95_ms_q25": 29.385052249999998,
    "steady_tpot_p95_ms_q75": 30.03668225,
    "decode_tokens_per_s": 35.504392032945084,
    "decode_tokens_per_s_q25": 35.41831321350345,
    "decode_tokens_per_s_q75": 35.65444674854093,
    "decode_ms": 422.482942,
    "decode_ms_q25": 420.71233899999993,
    "decode_ms_q75": 423.51222400000006,
    "end_to_end_ms": 2248.315806,
    "end_to_end_ms_q25": 2243.5500485000002,
    "end_to_end_ms_q75": 2250.3336799999997,
    "output_tokens_per_s_e2e": 7.116437983178952,
    "output_tokens_per_s_e2e_q25": 7.110062390899245,
    "output_tokens_per_s_e2e_q75": 7.131586927107119,
    "all_tokens_per_s_e2e": 910.9040618469058,
    "all_tokens_per_s_e2e_q25": 910.0879860351033,
    "all_tokens_per_s_e2e_q75": 912.8431266697112,
    "energy_net_psys_j": 45.893829260896986,
    "energy_net_psys_j_q25": 45.75310986901263,
    "energy_net_psys_j_q75": 46.46730132461012,
    "prefill_energy_net_psys_j": 31.261568620489392,
    "prefill_energy_net_psys_j_q25": 31.145657840992428,
    "prefill_energy_net_psys_j_q75": 31.623568567029025,
    "decode_energy_net_psys_j": 14.632260640407594,
    "decode_energy_net_psys_j_q25": 14.607452028020205,
    "decode_energy_net_psys_j_q75": 14.843732757581087,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.49741003110663534,
    "energy_reduction_vs_gpu": -0.2778917686193212
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 4096,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1603.065692,
    "ttft_ms_q25": 1546.1392455,
    "ttft_ms_q75": 1605.3625309999998,
    "prefill_tokens_per_s": 2545.123397226319,
    "prefill_tokens_per_s_q25": 2541.487217331717,
    "prefill_tokens_per_s_q75": 2642.412939060362,
    "handoff_plus_first_decode_ms": 43.987549,
    "handoff_plus_first_decode_ms_q25": 43.735888,
    "handoff_plus_first_decode_ms_q75": 50.556247,
    "steady_tpot_ms": 25.1619285,
    "steady_tpot_ms_q25": 25.134072500000002,
    "steady_tpot_ms_q75": 25.167222,
    "steady_tpot_p95_ms": 27.58854755,
    "steady_tpot_p95_ms_q25": 27.237907925,
    "steady_tpot_p95_ms_q75": 27.624445225000002,
    "decode_tokens_per_s": 37.30816477446453,
    "decode_tokens_per_s_q25": 36.65430131309493,
    "decode_tokens_per_s_q75": 37.38747916500209,
    "decode_ms": 402.05676400000016,
    "decode_ms_q25": 401.20564000000013,
    "decode_ms_q75": 409.3591815000001,
    "end_to_end_ms": 2003.420208,
    "end_to_end_ms_q25": 1954.647303,
    "end_to_end_ms_q75": 2006.568171,
    "output_tokens_per_s_e2e": 7.986342523704842,
    "output_tokens_per_s_e2e_q25": 7.973832940712187,
    "output_tokens_per_s_e2e_q75": 8.190719646150514,
    "all_tokens_per_s_e2e": 2044.5036860684395,
    "all_tokens_per_s_e2e_q25": 2041.30123282232,
    "all_tokens_per_s_e2e_q75": 2096.8242294145316,
    "energy_net_psys_j": 65.44541461298002,
    "energy_net_psys_j_q25": 64.21131973265548,
    "energy_net_psys_j_q75": 65.50187665354747,
    "prefill_energy_net_psys_j": 51.26185070605367,
    "prefill_energy_net_psys_j_q25": 49.51459644977226,
    "prefill_energy_net_psys_j_q75": 51.29657245381291,
    "decode_energy_net_psys_j": 14.227044492542767,
    "decode_energy_net_psys_j_q25": 14.205304199734556,
    "decode_energy_net_psys_j_q75": 14.71846357569143,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 4096,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1603.065692,
    "ttft_ms_q25": 1546.1392455,
    "ttft_ms_q75": 1605.3625309999998,
    "prefill_tokens_per_s": 2545.123397226319,
    "prefill_tokens_per_s_q25": 2541.487217331717,
    "prefill_tokens_per_s_q75": 2642.412939060362,
    "handoff_plus_first_decode_ms": 582.939636,
    "handoff_plus_first_decode_ms_q25": 580.6943795,
    "handoff_plus_first_decode_ms_q75": 628.354243,
    "steady_tpot_ms": 509.05022850000006,
    "steady_tpot_ms_q25": 509.02136825,
    "steady_tpot_ms_q75": 509.18104575,
    "steady_tpot_p95_ms": 509.9949101999999,
    "steady_tpot_p95_ms_q25": 509.78006782499995,
    "steady_tpot_p95_ms_q75": 510.52839697499996,
    "decode_tokens_per_s": 1.9455825050764004,
    "decode_tokens_per_s_q25": 1.9337595012668871,
    "decode_tokens_per_s_q75": 1.945695054301694,
    "decode_ms": 7709.773274000001,
    "decode_ms_q25": 7709.3273260000005,
    "decode_ms_q75": 7757.2007925,
    "end_to_end_ms": 9311.94707,
    "end_to_end_ms_q25": 9302.89409,
    "end_to_end_ms_q75": 9314.689857000001,
    "output_tokens_per_s_e2e": 1.7182228249070148,
    "output_tokens_per_s_e2e_q25": 1.7177170290072943,
    "output_tokens_per_s_e2e_q75": 1.7198965179135255,
    "all_tokens_per_s_e2e": 439.8650431761958,
    "all_tokens_per_s_e2e_q25": 439.73555942586734,
    "all_tokens_per_s_e2e_q75": 440.2935085858625,
    "energy_net_psys_j": 151.66995564329582,
    "energy_net_psys_j_q25": 151.24584492043118,
    "energy_net_psys_j_q75": 151.69562814205975,
    "prefill_energy_net_psys_j": 51.26185070605367,
    "prefill_energy_net_psys_j_q25": 49.51459644977226,
    "prefill_energy_net_psys_j_q75": 51.29657245381291,
    "decode_energy_net_psys_j": 100.39000643925154,
    "decode_energy_net_psys_j_q25": 99.97494496538219,
    "decode_energy_net_psys_j_q75": 102.14630994452824,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.21514514557909745,
    "energy_reduction_vs_gpu": -1.3175031671846815
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 4096,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 4026.3601529999996,
    "ttft_ms_q25": 4012.7952075,
    "ttft_ms_q75": 4027.595775,
    "prefill_tokens_per_s": 1013.3221681523034,
    "prefill_tokens_per_s_q25": 1013.0113874219435,
    "prefill_tokens_per_s_q75": 1016.7592445606155,
    "handoff_plus_first_decode_ms": 582.939636,
    "handoff_plus_first_decode_ms_q25": 580.6943795,
    "handoff_plus_first_decode_ms_q75": 628.354243,
    "steady_tpot_ms": 509.05022850000006,
    "steady_tpot_ms_q25": 509.02136825,
    "steady_tpot_ms_q75": 509.18104575,
    "steady_tpot_p95_ms": 509.9949101999999,
    "steady_tpot_p95_ms_q25": 509.78006782499995,
    "steady_tpot_p95_ms_q75": 510.52839697499996,
    "decode_tokens_per_s": 1.9455825050764004,
    "decode_tokens_per_s_q25": 1.9337595012668871,
    "decode_tokens_per_s_q75": 1.945695054301694,
    "decode_ms": 7709.773274000001,
    "decode_ms_q25": 7709.3273260000005,
    "decode_ms_q75": 7757.2007925,
    "end_to_end_ms": 11738.604671000001,
    "end_to_end_ms_q25": 11736.923101,
    "end_to_end_ms_q75": 11771.231622,
    "output_tokens_per_s_e2e": 1.363024009107973,
    "output_tokens_per_s_e2e_q25": 1.3592564853544513,
    "output_tokens_per_s_e2e_q75": 1.3632193199822265,
    "all_tokens_per_s_e2e": 348.9341463316411,
    "all_tokens_per_s_e2e_q25": 347.96966025073954,
    "all_tokens_per_s_e2e_q75": 348.98414591545,
    "energy_net_psys_j": 166.0139304845643,
    "energy_net_psys_j_q25": 165.5819915875751,
    "energy_net_psys_j_q75": 170.67634794726177,
    "prefill_energy_net_psys_j": 66.45404699305145,
    "prefill_energy_net_psys_j_q25": 65.6070466221929,
    "prefill_energy_net_psys_j_q75": 68.94509947660288,
    "decode_energy_net_psys_j": 100.39000643925154,
    "decode_energy_net_psys_j_q25": 99.97494496538219,
    "decode_energy_net_psys_j_q75": 102.14630994452824,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.17066936523975557,
    "energy_reduction_vs_gpu": -1.5366778018950495
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 4096,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 4026.3601529999996,
    "ttft_ms_q25": 4012.7952075,
    "ttft_ms_q75": 4027.595775,
    "prefill_tokens_per_s": 1013.3221681523034,
    "prefill_tokens_per_s_q25": 1013.0113874219435,
    "prefill_tokens_per_s_q75": 1016.7592445606155,
    "handoff_plus_first_decode_ms": 43.987549,
    "handoff_plus_first_decode_ms_q25": 43.735888,
    "handoff_plus_first_decode_ms_q75": 50.556247,
    "steady_tpot_ms": 25.1619285,
    "steady_tpot_ms_q25": 25.134072500000002,
    "steady_tpot_ms_q75": 25.167222,
    "steady_tpot_p95_ms": 27.58854755,
    "steady_tpot_p95_ms_q25": 27.237907925,
    "steady_tpot_p95_ms_q75": 27.624445225000002,
    "decode_tokens_per_s": 37.30816477446453,
    "decode_tokens_per_s_q25": 36.65430131309493,
    "decode_tokens_per_s_q75": 37.38747916500209,
    "decode_ms": 402.05676400000016,
    "decode_ms_q25": 401.20564000000013,
    "decode_ms_q75": 409.3591815000001,
    "end_to_end_ms": 4426.714669,
    "end_to_end_ms_q25": 4421.3032650000005,
    "end_to_end_ms_q75": 4428.801415,
    "output_tokens_per_s_e2e": 3.614418636928867,
    "output_tokens_per_s_e2e_q25": 3.6127164107929595,
    "output_tokens_per_s_e2e_q75": 3.6188478847649135,
    "all_tokens_per_s_e2e": 925.29117105379,
    "all_tokens_per_s_e2e_q25": 924.8554011629976,
    "all_tokens_per_s_e2e_q75": 926.4250584998179,
    "energy_net_psys_j": 80.6376108999778,
    "energy_net_psys_j_q25": 79.81235082192745,
    "energy_net_psys_j_q75": 83.64182275948609,
    "prefill_energy_net_psys_j": 66.45404699305145,
    "prefill_energy_net_psys_j_q25": 65.6070466221929,
    "prefill_energy_net_psys_j_q75": 68.94509947660288,
    "decode_energy_net_psys_j": 14.227044492542767,
    "decode_energy_net_psys_j_q25": 14.205304199734556,
    "decode_energy_net_psys_j_q75": 14.71846357569143,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.4525749585871942,
    "energy_reduction_vs_gpu": -0.2321353814753686
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 8192,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 4140.387078,
    "ttft_ms_q25": 4019.2779899999996,
    "ttft_ms_q75": 4158.200718,
    "prefill_tokens_per_s": 1974.6945988319019,
    "prefill_tokens_per_s_q25": 1966.2711367271445,
    "prefill_tokens_per_s_q75": 2036.0448040065307,
    "handoff_plus_first_decode_ms": 61.63259,
    "handoff_plus_first_decode_ms_q25": 60.214765,
    "handoff_plus_first_decode_ms_q75": 65.793271,
    "steady_tpot_ms": 29.060207000000002,
    "steady_tpot_ms_q25": 28.982426500000003,
    "steady_tpot_ms_q75": 30.7737385,
    "steady_tpot_p95_ms": 32.368322549999995,
    "steady_tpot_p95_ms_q25": 31.973692224999994,
    "steady_tpot_p95_ms_q75": 34.5626762,
    "decode_tokens_per_s": 31.061302329504077,
    "decode_tokens_per_s_q25": 30.204059251599833,
    "decode_tokens_per_s_q75": 31.43984592713797,
    "decode_ms": 482.91600400000016,
    "decode_ms_q25": 477.1707490000001,
    "decode_ms_q75": 497.02235200000064,
    "end_to_end_ms": 4611.812572,
    "end_to_end_ms_q25": 4496.4487389999995,
    "end_to_end_ms_q75": 4649.477815,
    "output_tokens_per_s_e2e": 3.4693517462400467,
    "output_tokens_per_s_e2e_q25": 3.4414725082996616,
    "output_tokens_per_s_e2e_q75": 3.5607075916662243,
    "all_tokens_per_s_e2e": 1776.308094074904,
    "all_tokens_per_s_e2e_q25": 1762.0339242494267,
    "all_tokens_per_s_e2e_q75": 1823.0822869331068,
    "energy_net_psys_j": 147.17802031892046,
    "energy_net_psys_j_q25": 145.08884473365845,
    "energy_net_psys_j_q75": 159.2070540992779,
    "prefill_energy_net_psys_j": 130.81313758436497,
    "prefill_energy_net_psys_j_q25": 128.31996768710655,
    "prefill_energy_net_psys_j_q75": 141.7355192197505,
    "decode_energy_net_psys_j": 17.172871358548306,
    "decode_energy_net_psys_j_q25": 16.7688770465519,
    "decode_energy_net_psys_j_q75": 17.875529191523803,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 8192,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 4140.387078,
    "ttft_ms_q25": 4019.2779899999996,
    "ttft_ms_q75": 4158.200718,
    "prefill_tokens_per_s": 1974.6945988319019,
    "prefill_tokens_per_s_q25": 1966.2711367271445,
    "prefill_tokens_per_s_q75": 2036.0448040065307,
    "handoff_plus_first_decode_ms": 611.2510540000001,
    "handoff_plus_first_decode_ms_q25": 596.9570105,
    "handoff_plus_first_decode_ms_q75": 660.72153,
    "steady_tpot_ms": 515.2153075,
    "steady_tpot_ms_q25": 515.1769075,
    "steady_tpot_ms_q75": 515.3487475,
    "steady_tpot_p95_ms": 515.8123436500001,
    "steady_tpot_p95_ms_q25": 515.7713989000001,
    "steady_tpot_p95_ms_q75": 516.6305286250001,
    "decode_tokens_per_s": 1.9172556261856764,
    "decode_tokens_per_s_q25": 1.9042417474267752,
    "decode_tokens_per_s_q75": 1.9251361950512125,
    "decode_ms": 7823.682869999999,
    "decode_ms_q25": 7791.787093,
    "decode_ms_q75": 7877.519037499999,
    "end_to_end_ms": 11935.905674000001,
    "end_to_end_ms_q25": 11882.7148905,
    "end_to_end_ms_q75": 11949.987810999999,
    "output_tokens_per_s_e2e": 1.3404931671714548,
    "output_tokens_per_s_e2e_q25": 1.3389153589080698,
    "output_tokens_per_s_e2e_q75": 1.3465206185882077,
    "all_tokens_per_s_e2e": 686.3325015917849,
    "all_tokens_per_s_e2e_q25": 685.5246637609317,
    "all_tokens_per_s_e2e_q75": 689.4185567171623,
    "energy_net_psys_j": 231.65250603027113,
    "energy_net_psys_j_q25": 231.2544132075988,
    "energy_net_psys_j_q75": 245.3501960925922,
    "prefill_energy_net_psys_j": 130.81313758436497,
    "prefill_energy_net_psys_j_q25": 128.31996768710655,
    "prefill_energy_net_psys_j_q75": 141.7355192197505,
    "decode_energy_net_psys_j": 105.0295225950783,
    "decode_energy_net_psys_j_q25": 102.93444552049223,
    "decode_energy_net_psys_j_q75": 105.70975394742779,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.3863814525650883,
    "energy_reduction_vs_gpu": -0.5739612853081097
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 8192,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 8989.586643999999,
    "ttft_ms_q25": 8978.298316999999,
    "ttft_ms_q75": 9025.349912999998,
    "prefill_tokens_per_s": 909.4967681808797,
    "prefill_tokens_per_s_q25": 905.9070792627788,
    "prefill_tokens_per_s_q75": 910.6417091295481,
    "handoff_plus_first_decode_ms": 611.2510540000001,
    "handoff_plus_first_decode_ms_q25": 596.9570105,
    "handoff_plus_first_decode_ms_q75": 660.72153,
    "steady_tpot_ms": 515.2153075,
    "steady_tpot_ms_q25": 515.1769075,
    "steady_tpot_ms_q75": 515.3487475,
    "steady_tpot_p95_ms": 515.8123436500001,
    "steady_tpot_p95_ms_q25": 515.7713989000001,
    "steady_tpot_p95_ms_q75": 516.6305286250001,
    "decode_tokens_per_s": 1.9172556261856764,
    "decode_tokens_per_s_q25": 1.9042417474267752,
    "decode_tokens_per_s_q75": 1.9251361950512125,
    "decode_ms": 7823.682869999999,
    "decode_ms_q25": 7791.787093,
    "decode_ms_q75": 7877.519037499999,
    "end_to_end_ms": 16884.796051999998,
    "end_to_end_ms_q25": 16805.848679,
    "end_to_end_ms_q75": 16902.8689505,
    "output_tokens_per_s_e2e": 0.9475980610440838,
    "output_tokens_per_s_e2e_q25": 0.9465859516869999,
    "output_tokens_per_s_e2e_q75": 0.9520705198821406,
    "all_tokens_per_s_e2e": 485.1702072545709,
    "all_tokens_per_s_e2e_q25": 484.65200726374394,
    "all_tokens_per_s_e2e_q75": 487.460106179656,
    "energy_net_psys_j": 262.7768533889938,
    "energy_net_psys_j_q25": 262.45406954808317,
    "energy_net_psys_j_q75": 267.7749618664778,
    "prefill_energy_net_psys_j": 161.29191726126638,
    "prefill_energy_net_psys_j_q25": 159.51962402759094,
    "prefill_energy_net_psys_j_q75": 163.83750115272545,
    "decode_energy_net_psys_j": 105.0295225950783,
    "decode_energy_net_psys_j_q25": 102.93444552049223,
    "decode_energy_net_psys_j_q75": 105.70975394742779,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.27313404069537056,
    "energy_reduction_vs_gpu": -0.7854354394737877
  },
  {
    "precision": "INT8",
    "model": "qwen25_1_5b",
    "model_label": "Qwen2.5 1.5B",
    "family": "Qwen2.5",
    "size_b": 1.5,
    "context_length": 8192,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 8989.586643999999,
    "ttft_ms_q25": 8978.298316999999,
    "ttft_ms_q75": 9025.349912999998,
    "prefill_tokens_per_s": 909.4967681808797,
    "prefill_tokens_per_s_q25": 905.9070792627788,
    "prefill_tokens_per_s_q75": 910.6417091295481,
    "handoff_plus_first_decode_ms": 61.63259,
    "handoff_plus_first_decode_ms_q25": 60.214765,
    "handoff_plus_first_decode_ms_q75": 65.793271,
    "steady_tpot_ms": 29.060207000000002,
    "steady_tpot_ms_q25": 28.982426500000003,
    "steady_tpot_ms_q75": 30.7737385,
    "steady_tpot_p95_ms": 32.368322549999995,
    "steady_tpot_p95_ms_q25": 31.973692224999994,
    "steady_tpot_p95_ms_q75": 34.5626762,
    "decode_tokens_per_s": 31.061302329504077,
    "decode_tokens_per_s_q25": 30.204059251599833,
    "decode_tokens_per_s_q75": 31.43984592713797,
    "decode_ms": 482.91600400000016,
    "decode_ms_q25": 477.1707490000001,
    "decode_ms_q75": 497.02235200000064,
    "end_to_end_ms": 9478.13869,
    "end_to_end_ms_q25": 9475.320669,
    "end_to_end_ms_q75": 9505.338683,
    "output_tokens_per_s_e2e": 1.6880951548937506,
    "output_tokens_per_s_e2e_q25": 1.6832783712480628,
    "output_tokens_per_s_e2e_q75": 1.6885973545524116,
    "all_tokens_per_s_e2e": 864.3047193056003,
    "all_tokens_per_s_e2e_q25": 861.8385260790081,
    "all_tokens_per_s_e2e_q75": 864.5618455308347,
    "energy_net_psys_j": 177.65679999582187,
    "energy_net_psys_j_q25": 176.28850107414283,
    "energy_net_psys_j_q75": 181.30903603225283,
    "prefill_energy_net_psys_j": 161.29191726126638,
    "prefill_energy_net_psys_j_q25": 159.51962402759094,
    "prefill_energy_net_psys_j_q75": 163.83750115272545,
    "decode_energy_net_psys_j": 17.172871358548306,
    "decode_energy_net_psys_j_q25": 16.7688770465519,
    "decode_energy_net_psys_j_q75": 17.875529191523803,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.48657365362945537,
    "energy_reduction_vs_gpu": -0.20708784919688994
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 489.558879,
    "ttft_ms_q25": 472.72711449999997,
    "ttft_ms_q75": 495.3317935,
    "prefill_tokens_per_s": 2058.9964624050867,
    "prefill_tokens_per_s_q25": 2035.2760495981352,
    "prefill_tokens_per_s_q75": 2135.0151003699957,
    "handoff_plus_first_decode_ms": 52.612127,
    "handoff_plus_first_decode_ms_q25": 52.0652285,
    "handoff_plus_first_decode_ms_q75": 59.6678235,
    "steady_tpot_ms": 43.007611000000004,
    "steady_tpot_ms_q25": 42.886047250000004,
    "steady_tpot_ms_q75": 43.379581,
    "steady_tpot_p95_ms": 45.50781405,
    "steady_tpot_p95_ms_q25": 45.48998077500001,
    "steady_tpot_p95_ms_q75": 45.900485325,
    "decode_tokens_per_s": 22.38918474131744,
    "decode_tokens_per_s_q25": 22.30778113768578,
    "decode_tokens_per_s_q75": 22.55167896479459,
    "decode_ms": 669.9663330000003,
    "decode_ms_q25": 665.1734815000002,
    "decode_ms_q75": 672.4200695000001,
    "end_to_end_ms": 1159.5252120000002,
    "end_to_end_ms_q25": 1145.147184,
    "end_to_end_ms_q75": 1160.505275,
    "output_tokens_per_s_e2e": 13.798751277173608,
    "output_tokens_per_s_e2e_q25": 13.78710787098456,
    "output_tokens_per_s_e2e_q75": 13.974206038364706,
    "all_tokens_per_s_e2e": 883.1200817391109,
    "all_tokens_per_s_e2e_q25": 882.3749037430118,
    "all_tokens_per_s_e2e_q75": 894.3491864553412,
    "energy_net_psys_j": 39.81794016999228,
    "energy_net_psys_j_q25": 39.57505361206672,
    "energy_net_psys_j_q75": 39.917907369158414,
    "prefill_energy_net_psys_j": 16.575855736145257,
    "prefill_energy_net_psys_j_q25": 16.144735711557786,
    "prefill_energy_net_psys_j_q75": 16.611346368933763,
    "decode_energy_net_psys_j": 23.171103168270015,
    "decode_energy_net_psys_j_q25": 22.963707243132955,
    "decode_energy_net_psys_j_q75": 23.73768102481212,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 489.558879,
    "ttft_ms_q25": 472.72711449999997,
    "ttft_ms_q75": 495.3317935,
    "prefill_tokens_per_s": 2058.9964624050867,
    "prefill_tokens_per_s_q25": 2035.2760495981352,
    "prefill_tokens_per_s_q75": 2135.0151003699957,
    "handoff_plus_first_decode_ms": 1343.034015,
    "handoff_plus_first_decode_ms_q25": 1335.152375,
    "handoff_plus_first_decode_ms_q75": 1489.3587885000002,
    "steady_tpot_ms": 1336.6938610000002,
    "steady_tpot_ms_q25": 1330.15157425,
    "steady_tpot_ms_q75": 1336.81788,
    "steady_tpot_p95_ms": 1338.9391349,
    "steady_tpot_p95_ms_q25": 1338.3699532,
    "steady_tpot_p95_ms_q75": 1340.4359989499999,
    "decode_tokens_per_s": 0.7489140064645603,
    "decode_tokens_per_s_q25": 0.7435102330863657,
    "decode_tokens_per_s_q75": 0.7512735981285256,
    "decode_ms": 20029.001822000002,
    "decode_ms_q25": 19966.291917,
    "decode_ms_q75": 20175.636749,
    "end_to_end_ms": 20518.560701000002,
    "end_to_end_ms_q25": 20461.623710499996,
    "end_to_end_ms_q75": 20648.363863500002,
    "output_tokens_per_s_e2e": 0.779781790406976,
    "output_tokens_per_s_e2e_q25": 0.7749104202650343,
    "output_tokens_per_s_e2e_q75": 0.7819576840632585,
    "all_tokens_per_s_e2e": 49.90603458604647,
    "all_tokens_per_s_e2e_q25": 49.594266896962196,
    "all_tokens_per_s_e2e_q75": 50.045291780048544,
    "energy_net_psys_j": 247.05295659878448,
    "energy_net_psys_j_q25": 239.63122704771592,
    "energy_net_psys_j_q75": 248.3044201335482,
    "prefill_energy_net_psys_j": 16.575855736145257,
    "prefill_energy_net_psys_j_q25": 16.144735711557786,
    "prefill_energy_net_psys_j_q75": 16.611346368933763,
    "decode_energy_net_psys_j": 230.47710086263922,
    "decode_energy_net_psys_j_q25": 223.01988067878216,
    "decode_energy_net_psys_j_q75": 232.15968442199042,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.056511040364711794,
    "energy_reduction_vs_gpu": -5.2045639614720525
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1940.663878,
    "ttft_ms_q25": 1939.8918695,
    "ttft_ms_q75": 1941.3668505,
    "prefill_tokens_per_s": 519.4098841262609,
    "prefill_tokens_per_s_q25": 519.2218729330253,
    "prefill_tokens_per_s_q75": 519.6166732235179,
    "handoff_plus_first_decode_ms": 1343.034015,
    "handoff_plus_first_decode_ms_q25": 1335.152375,
    "handoff_plus_first_decode_ms_q75": 1489.3587885000002,
    "steady_tpot_ms": 1336.6938610000002,
    "steady_tpot_ms_q25": 1330.15157425,
    "steady_tpot_ms_q75": 1336.81788,
    "steady_tpot_p95_ms": 1338.9391349,
    "steady_tpot_p95_ms_q25": 1338.3699532,
    "steady_tpot_p95_ms_q75": 1340.4359989499999,
    "decode_tokens_per_s": 0.7489140064645603,
    "decode_tokens_per_s_q25": 0.7435102330863657,
    "decode_tokens_per_s_q75": 0.7512735981285256,
    "decode_ms": 20029.001822000002,
    "decode_ms_q25": 19966.291917,
    "decode_ms_q75": 20175.636749,
    "end_to_end_ms": 21968.121683,
    "end_to_end_ms_q25": 21906.183786499998,
    "end_to_end_ms_q75": 22116.231591000003,
    "output_tokens_per_s_e2e": 0.7283280851626736,
    "output_tokens_per_s_e2e_q25": 0.7234830011213556,
    "output_tokens_per_s_e2e_q75": 0.7303932106782905,
    "all_tokens_per_s_e2e": 46.61299745041111,
    "all_tokens_per_s_e2e_q25": 46.302912071766755,
    "all_tokens_per_s_e2e_q75": 46.745165483410595,
    "energy_net_psys_j": 263.66322710277404,
    "energy_net_psys_j_q25": 256.05902519237173,
    "energy_net_psys_j_q75": 265.27150792736074,
    "prefill_energy_net_psys_j": 33.037520770605816,
    "prefill_energy_net_psys_j_q25": 32.964841778825075,
    "prefill_energy_net_psys_j_q75": 33.111823505370324,
    "decode_energy_net_psys_j": 230.47710086263922,
    "decode_energy_net_psys_j_q25": 223.01988067878216,
    "decode_energy_net_psys_j_q75": 232.15968442199042,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.05278217358461271,
    "energy_reduction_vs_gpu": -5.621719405301551
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 1024,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1940.663878,
    "ttft_ms_q25": 1939.8918695,
    "ttft_ms_q75": 1941.3668505,
    "prefill_tokens_per_s": 519.4098841262609,
    "prefill_tokens_per_s_q25": 519.2218729330253,
    "prefill_tokens_per_s_q75": 519.6166732235179,
    "handoff_plus_first_decode_ms": 52.612127,
    "handoff_plus_first_decode_ms_q25": 52.0652285,
    "handoff_plus_first_decode_ms_q75": 59.6678235,
    "steady_tpot_ms": 43.007611000000004,
    "steady_tpot_ms_q25": 42.886047250000004,
    "steady_tpot_ms_q75": 43.379581,
    "steady_tpot_p95_ms": 45.50781405,
    "steady_tpot_p95_ms_q25": 45.48998077500001,
    "steady_tpot_p95_ms_q75": 45.900485325,
    "decode_tokens_per_s": 22.38918474131744,
    "decode_tokens_per_s_q25": 22.30778113768578,
    "decode_tokens_per_s_q75": 22.55167896479459,
    "decode_ms": 669.9663330000003,
    "decode_ms_q25": 665.1734815000002,
    "decode_ms_q75": 672.4200695000001,
    "end_to_end_ms": 2609.0861940000004,
    "end_to_end_ms_q25": 2605.065351,
    "end_to_end_ms_q75": 2613.0149115000004,
    "output_tokens_per_s_e2e": 6.132415263548781,
    "output_tokens_per_s_e2e_q25": 6.123208902778842,
    "output_tokens_per_s_e2e_q75": 6.141895100940912,
    "all_tokens_per_s_e2e": 392.474576867122,
    "all_tokens_per_s_e2e_q25": 391.8853697778459,
    "all_tokens_per_s_e2e_q75": 393.0812864602184,
    "energy_net_psys_j": 56.06326595531434,
    "energy_net_psys_j_q25": 56.002851756722535,
    "energy_net_psys_j_q75": 56.702522803637194,
    "prefill_energy_net_psys_j": 33.037520770605816,
    "prefill_energy_net_psys_j_q25": 32.964841778825075,
    "prefill_energy_net_psys_j_q75": 33.111823505370324,
    "decode_energy_net_psys_j": 23.171103168270015,
    "decode_energy_net_psys_j_q25": 22.963707243132955,
    "decode_energy_net_psys_j_q75": 23.73768102481212,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.4444181317836524,
    "energy_reduction_vs_gpu": -0.4079901098843106
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1066.727625,
    "ttft_ms_q25": 1033.1412235,
    "ttft_ms_q75": 1073.7445535,
    "prefill_tokens_per_s": 1904.8911384478301,
    "prefill_tokens_per_s_q25": 1892.5234836586933,
    "prefill_tokens_per_s_q75": 1968.8980715159805,
    "handoff_plus_first_decode_ms": 62.479971,
    "handoff_plus_first_decode_ms_q25": 61.4363465,
    "handoff_plus_first_decode_ms_q75": 69.4866055,
    "steady_tpot_ms": 49.103414,
    "steady_tpot_ms_q25": 48.90559575,
    "steady_tpot_ms_q75": 49.2539075,
    "steady_tpot_p95_ms": 51.85995685,
    "steady_tpot_p95_ms_q25": 51.31604597500001,
    "steady_tpot_p95_ms_q75": 52.1106489,
    "decode_tokens_per_s": 19.680266234011846,
    "decode_tokens_per_s_q25": 19.67672433241007,
    "decode_tokens_per_s_q75": 19.762366803479843,
    "decode_ms": 762.1848110000001,
    "decode_ms_q25": 759.0314985,
    "decode_ms_q75": 762.3220325,
    "end_to_end_ms": 1822.605811,
    "end_to_end_ms_q25": 1792.3099435,
    "end_to_end_ms_q75": 1832.776052,
    "output_tokens_per_s_e2e": 8.778639848196995,
    "output_tokens_per_s_e2e_q25": 8.730195201032029,
    "output_tokens_per_s_e2e_q75": 8.929578772745742,
    "all_tokens_per_s_e2e": 1123.6659005692154,
    "all_tokens_per_s_e2e_q25": 1117.4649857320996,
    "all_tokens_per_s_e2e_q75": 1142.986082911455,
    "energy_net_psys_j": 59.92964746310727,
    "energy_net_psys_j_q25": 58.32317898069759,
    "energy_net_psys_j_q75": 60.48587625661966,
    "prefill_energy_net_psys_j": 36.1566927687641,
    "prefill_energy_net_psys_j_q25": 34.25054310827647,
    "prefill_energy_net_psys_j_q75": 36.174813476497434,
    "decode_energy_net_psys_j": 24.372317050499085,
    "decode_energy_net_psys_j_q25": 24.054515164687796,
    "decode_energy_net_psys_j_q75": 24.62886466593352,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1066.727625,
    "ttft_ms_q25": 1033.1412235,
    "ttft_ms_q75": 1073.7445535,
    "prefill_tokens_per_s": 1904.8911384478301,
    "prefill_tokens_per_s_q25": 1892.5234836586933,
    "prefill_tokens_per_s_q75": 1968.8980715159805,
    "handoff_plus_first_decode_ms": 1375.780532,
    "handoff_plus_first_decode_ms_q25": 1375.4221160000002,
    "handoff_plus_first_decode_ms_q75": 1451.828402,
    "steady_tpot_ms": 1336.4961575,
    "steady_tpot_ms_q25": 1336.03517975,
    "steady_tpot_ms_q75": 1337.2230605,
    "steady_tpot_p95_ms": 1337.7854578,
    "steady_tpot_p95_ms_q25": 1337.550065375,
    "steady_tpot_p95_ms_q75": 1340.5493260499998,
    "decode_tokens_per_s": 0.7468572382967064,
    "decode_tokens_per_s_q25": 0.743551630479671,
    "decode_tokens_per_s_q75": 0.7475004271214611,
    "decode_ms": 20084.159637,
    "decode_ms_q25": 20066.893023,
    "decode_ms_q75": 20173.8465155,
    "end_to_end_ms": 21164.921119,
    "end_to_end_ms_q25": 21140.637576499998,
    "end_to_end_ms_q75": 21214.0046675,
    "output_tokens_per_s_e2e": 0.7559678540751381,
    "output_tokens_per_s_e2e_q25": 0.7542227837235032,
    "output_tokens_per_s_e2e_q75": 0.7568372076332148,
    "all_tokens_per_s_e2e": 96.76388532161768,
    "all_tokens_per_s_e2e_q25": 96.5405163166084,
    "all_tokens_per_s_e2e_q75": 96.8751625770515,
    "energy_net_psys_j": 270.77864527522553,
    "energy_net_psys_j_q25": 270.6503646776142,
    "energy_net_psys_j_q75": 272.11536314097793,
    "prefill_energy_net_psys_j": 36.1566927687641,
    "prefill_energy_net_psys_j_q25": 34.25054310827647,
    "prefill_energy_net_psys_j_q75": 36.174813476497434,
    "decode_energy_net_psys_j": 234.6219525064614,
    "decode_energy_net_psys_j_q25": 234.47555120111676,
    "decode_energy_net_psys_j_q75": 237.8648200327014,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.08611446273540917,
    "energy_reduction_vs_gpu": -3.5182752900710295
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3904.601155,
    "ttft_ms_q25": 3901.976305,
    "ttft_ms_q75": 3908.4971225,
    "prefill_tokens_per_s": 520.4116680132468,
    "prefill_tokens_per_s_q25": 519.8934412148244,
    "prefill_tokens_per_s_q75": 520.7619833368458,
    "handoff_plus_first_decode_ms": 1375.780532,
    "handoff_plus_first_decode_ms_q25": 1375.4221160000002,
    "handoff_plus_first_decode_ms_q75": 1451.828402,
    "steady_tpot_ms": 1336.4961575,
    "steady_tpot_ms_q25": 1336.03517975,
    "steady_tpot_ms_q75": 1337.2230605,
    "steady_tpot_p95_ms": 1337.7854578,
    "steady_tpot_p95_ms_q25": 1337.550065375,
    "steady_tpot_p95_ms_q75": 1340.5493260499998,
    "decode_tokens_per_s": 0.7468572382967064,
    "decode_tokens_per_s_q25": 0.743551630479671,
    "decode_tokens_per_s_q75": 0.7475004271214611,
    "decode_ms": 20084.159637,
    "decode_ms_q25": 20066.893023,
    "decode_ms_q75": 20173.8465155,
    "end_to_end_ms": 23996.552727000002,
    "end_to_end_ms_q25": 23975.3901455,
    "end_to_end_ms_q75": 24079.718788000002,
    "output_tokens_per_s_e2e": 0.666762438006248,
    "output_tokens_per_s_e2e_q25": 0.6644675131271092,
    "output_tokens_per_s_e2e_q75": 0.667351495382352,
    "all_tokens_per_s_e2e": 85.34559206479975,
    "all_tokens_per_s_e2e_q25": 85.05184168026997,
    "all_tokens_per_s_e2e_q75": 85.42099140894105,
    "energy_net_psys_j": 303.472746758231,
    "energy_net_psys_j_q25": 303.4300304842504,
    "energy_net_psys_j_q75": 306.82081674932994,
    "prefill_energy_net_psys_j": 69.05816431449767,
    "prefill_energy_net_psys_j_q25": 68.95447928313362,
    "prefill_energy_net_psys_j_q75": 69.05968174799257,
    "decode_energy_net_psys_j": 234.6219525064614,
    "decode_energy_net_psys_j_q25": 234.47555120111676,
    "decode_energy_net_psys_j_q75": 237.8648200327014,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.07595281837916967,
    "energy_reduction_vs_gpu": -4.063816651767374
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 2048,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 3904.601155,
    "ttft_ms_q25": 3901.976305,
    "ttft_ms_q75": 3908.4971225,
    "prefill_tokens_per_s": 520.4116680132468,
    "prefill_tokens_per_s_q25": 519.8934412148244,
    "prefill_tokens_per_s_q75": 520.7619833368458,
    "handoff_plus_first_decode_ms": 62.479971,
    "handoff_plus_first_decode_ms_q25": 61.4363465,
    "handoff_plus_first_decode_ms_q75": 69.4866055,
    "steady_tpot_ms": 49.103414,
    "steady_tpot_ms_q25": 48.90559575,
    "steady_tpot_ms_q75": 49.2539075,
    "steady_tpot_p95_ms": 51.85995685,
    "steady_tpot_p95_ms_q25": 51.31604597500001,
    "steady_tpot_p95_ms_q75": 52.1106489,
    "decode_tokens_per_s": 19.680266234011846,
    "decode_tokens_per_s_q25": 19.67672433241007,
    "decode_tokens_per_s_q75": 19.762366803479843,
    "decode_ms": 762.1848110000001,
    "decode_ms_q25": 759.0314985,
    "decode_ms_q75": 762.3220325,
    "end_to_end_ms": 4661.810708999999,
    "end_to_end_ms_q25": 4661.145025,
    "end_to_end_ms_q75": 4668.194305,
    "output_tokens_per_s_e2e": 3.432142787160087,
    "output_tokens_per_s_e2e_q25": 3.4274558585364656,
    "output_tokens_per_s_e2e_q75": 3.4326330205399516,
    "all_tokens_per_s_e2e": 439.31427675649115,
    "all_tokens_per_s_e2e_q25": 438.7143498926676,
    "all_tokens_per_s_e2e_q75": 439.3770266291138,
    "energy_net_psys_j": 93.43351623198654,
    "energy_net_psys_j_q25": 93.11419691268037,
    "energy_net_psys_j_q75": 93.58486138256204,
    "prefill_energy_net_psys_j": 69.05816431449767,
    "prefill_energy_net_psys_j_q25": 68.95447928313362,
    "prefill_energy_net_psys_j_q75": 69.05968174799257,
    "decode_energy_net_psys_j": 24.372317050499085,
    "decode_energy_net_psys_j_q25": 24.054515164687796,
    "decode_energy_net_psys_j_q75": 24.62886466593352,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.39096521175373194,
    "energy_reduction_vs_gpu": -0.5590533264775881
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2492.958318,
    "ttft_ms_q25": 2438.228104,
    "ttft_ms_q75": 2499.2266405,
    "prefill_tokens_per_s": 1636.6097942917954,
    "prefill_tokens_per_s_q25": 1632.515274786158,
    "prefill_tokens_per_s_q75": 1674.189858722264,
    "handoff_plus_first_decode_ms": 65.289453,
    "handoff_plus_first_decode_ms_q25": 63.5090885,
    "handoff_plus_first_decode_ms_q75": 70.8923875,
    "steady_tpot_ms": 46.4175805,
    "steady_tpot_ms_q25": 46.3737525,
    "steady_tpot_ms_q75": 46.54891725,
    "steady_tpot_p95_ms": 48.02061165,
    "steady_tpot_p95_ms_q25": 47.922808375,
    "steady_tpot_p95_ms_q75": 48.0672597,
    "decode_tokens_per_s": 20.91488730532362,
    "decode_tokens_per_s_q25": 20.683897259594175,
    "decode_tokens_per_s_q75": 20.93192171329561,
    "decode_ms": 717.192485,
    "decode_ms_q25": 716.609308,
    "decode_ms_q75": 725.2922785000001,
    "end_to_end_ms": 3210.150803,
    "end_to_end_ms_q25": 3163.5203825,
    "end_to_end_ms_q75": 3215.8359485,
    "output_tokens_per_s_e2e": 4.984189523136244,
    "output_tokens_per_s_e2e_q25": 4.9753937270474395,
    "output_tokens_per_s_e2e_q75": 5.058755791483557,
    "all_tokens_per_s_e2e": 1275.9525179228785,
    "all_tokens_per_s_e2e_q25": 1273.7007941241445,
    "all_tokens_per_s_e2e_q75": 1295.0414826197905,
    "energy_net_psys_j": 112.90325488563047,
    "energy_net_psys_j_q25": 110.92819247846842,
    "energy_net_psys_j_q75": 113.06037754803806,
    "prefill_energy_net_psys_j": 87.98267846640633,
    "prefill_energy_net_psys_j_q25": 85.82492177695964,
    "prefill_energy_net_psys_j_q75": 88.11796715350407,
    "decode_energy_net_psys_j": 24.96424436984384,
    "decode_energy_net_psys_j_q25": 24.942410394533994,
    "decode_energy_net_psys_j_q75": 25.125104676818623,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 2492.958318,
    "ttft_ms_q25": 2438.228104,
    "ttft_ms_q75": 2499.2266405,
    "prefill_tokens_per_s": 1636.6097942917954,
    "prefill_tokens_per_s_q25": 1632.515274786158,
    "prefill_tokens_per_s_q75": 1674.189858722264,
    "handoff_plus_first_decode_ms": 1415.829474,
    "handoff_plus_first_decode_ms_q25": 1414.501694,
    "handoff_plus_first_decode_ms_q75": 1490.9803935,
    "steady_tpot_ms": 1338.565505,
    "steady_tpot_ms_q25": 1336.10523575,
    "steady_tpot_ms_q75": 1339.1301762500002,
    "steady_tpot_p95_ms": 1341.0676509999998,
    "steady_tpot_p95_ms_q25": 1340.696435675,
    "steady_tpot_p95_ms_q75": 1341.1681983999997,
    "decode_tokens_per_s": 0.7437284279239131,
    "decode_tokens_per_s_q25": 0.7411313284209509,
    "decode_tokens_per_s_q75": 0.7459565858680587,
    "decode_ms": 20168.652208000003,
    "decode_ms_q25": 20108.5882425,
    "decode_ms_q75": 20239.576472,
    "end_to_end_ms": 22674.147171000004,
    "end_to_end_ms_q25": 22607.814883000003,
    "end_to_end_ms_q75": 22684.0728985,
    "output_tokens_per_s_e2e": 0.7056494729144138,
    "output_tokens_per_s_e2e_q25": 0.7053408413464786,
    "output_tokens_per_s_e2e_q75": 0.7077259709383189,
    "all_tokens_per_s_e2e": 180.64626506608994,
    "all_tokens_per_s_e2e_q25": 180.5672553846985,
    "all_tokens_per_s_e2e_q75": 181.17784856020964,
    "energy_net_psys_j": 326.3680963287806,
    "energy_net_psys_j_q25": 318.3110116269386,
    "energy_net_psys_j_q75": 328.34558197099636,
    "prefill_energy_net_psys_j": 87.98267846640633,
    "prefill_energy_net_psys_j_q25": 85.82492177695964,
    "prefill_energy_net_psys_j_q75": 88.11796715350407,
    "decode_energy_net_psys_j": 242.0698117726104,
    "decode_energy_net_psys_j_q25": 232.17053011565034,
    "decode_energy_net_psys_j_q75": 242.385371506939,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.14157757638204577,
    "energy_reduction_vs_gpu": -1.8906881086766476
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 8360.637991000001,
    "ttft_ms_q25": 8355.382377500002,
    "ttft_ms_q75": 8375.669299000001,
    "prefill_tokens_per_s": 488.00103585300656,
    "prefill_tokens_per_s_q25": 487.1268187622304,
    "prefill_tokens_per_s_q75": 488.3081862597645,
    "handoff_plus_first_decode_ms": 1415.829474,
    "handoff_plus_first_decode_ms_q25": 1414.501694,
    "handoff_plus_first_decode_ms_q75": 1490.9803935,
    "steady_tpot_ms": 1338.565505,
    "steady_tpot_ms_q25": 1336.10523575,
    "steady_tpot_ms_q75": 1339.1301762500002,
    "steady_tpot_p95_ms": 1341.0676509999998,
    "steady_tpot_p95_ms_q25": 1340.696435675,
    "steady_tpot_p95_ms_q75": 1341.1681983999997,
    "decode_tokens_per_s": 0.7437284279239131,
    "decode_tokens_per_s_q25": 0.7411313284209509,
    "decode_tokens_per_s_q75": 0.7459565858680587,
    "decode_ms": 20168.652208000003,
    "decode_ms_q25": 20108.5882425,
    "decode_ms_q75": 20239.576472,
    "end_to_end_ms": 28559.352815000002,
    "end_to_end_ms_q25": 28479.001928,
    "end_to_end_ms_q75": 28615.245771,
    "output_tokens_per_s_e2e": 0.5602367849034887,
    "output_tokens_per_s_e2e_q25": 0.5591446312131052,
    "output_tokens_per_s_e2e_q75": 0.5618219138073945,
    "all_tokens_per_s_e2e": 143.4206169352931,
    "all_tokens_per_s_e2e_q25": 143.14102559055493,
    "all_tokens_per_s_e2e_q75": 143.826409934693,
    "energy_net_psys_j": 388.6761570662737,
    "energy_net_psys_j_q25": 380.1838561448533,
    "energy_net_psys_j_q75": 389.2394782311085,
    "prefill_energy_net_psys_j": 147.73298762333283,
    "prefill_energy_net_psys_j_q25": 146.8541067241694,
    "prefill_energy_net_psys_j_q75": 148.5766471940377,
    "decode_energy_net_psys_j": 242.0698117726104,
    "decode_energy_net_psys_j_q25": 232.17053011565034,
    "decode_energy_net_psys_j_q75": 242.385371506939,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.11240278530800453,
    "energy_reduction_vs_gpu": -2.4425593616410577
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 4096,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 8360.637991000001,
    "ttft_ms_q25": 8355.382377500002,
    "ttft_ms_q75": 8375.669299000001,
    "prefill_tokens_per_s": 488.00103585300656,
    "prefill_tokens_per_s_q25": 487.1268187622304,
    "prefill_tokens_per_s_q75": 488.3081862597645,
    "handoff_plus_first_decode_ms": 65.289453,
    "handoff_plus_first_decode_ms_q25": 63.5090885,
    "handoff_plus_first_decode_ms_q75": 70.8923875,
    "steady_tpot_ms": 46.4175805,
    "steady_tpot_ms_q25": 46.3737525,
    "steady_tpot_ms_q75": 46.54891725,
    "steady_tpot_p95_ms": 48.02061165,
    "steady_tpot_p95_ms_q25": 47.922808375,
    "steady_tpot_p95_ms_q75": 48.0672597,
    "decode_tokens_per_s": 20.91488730532362,
    "decode_tokens_per_s_q25": 20.683897259594175,
    "decode_tokens_per_s_q75": 20.93192171329561,
    "decode_ms": 717.192485,
    "decode_ms_q25": 716.609308,
    "decode_ms_q75": 725.2922785000001,
    "end_to_end_ms": 9094.030063000002,
    "end_to_end_ms_q25": 9080.674656000001,
    "end_to_end_ms_q75": 9100.378400500002,
    "output_tokens_per_s_e2e": 1.7593959871649918,
    "output_tokens_per_s_e2e_q25": 1.7581695048246289,
    "output_tokens_per_s_e2e_q75": 1.761987431100899,
    "all_tokens_per_s_e2e": 450.4053727142379,
    "all_tokens_per_s_e2e_q25": 450.091393235105,
    "all_tokens_per_s_e2e_q75": 451.06878236183013,
    "energy_net_psys_j": 172.69723199317667,
    "energy_net_psys_j_q25": 171.97921140098805,
    "energy_net_psys_j_q75": 173.51905758857168,
    "prefill_energy_net_psys_j": 147.73298762333283,
    "prefill_energy_net_psys_j_q25": 146.8541067241694,
    "prefill_energy_net_psys_j_q75": 148.5766471940377,
    "decode_energy_net_psys_j": 24.96424436984384,
    "decode_energy_net_psys_j_q25": 24.942410394533994,
    "decode_energy_net_psys_j_q75": 25.125104676818623,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.35299540256204226,
    "energy_reduction_vs_gpu": -0.5296036608343728
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 6571.898811999999,
    "ttft_ms_q25": 6458.2163755,
    "ttft_ms_q75": 6574.1596835,
    "prefill_tokens_per_s": 1244.08488838431,
    "prefill_tokens_per_s_q25": 1243.657191273974,
    "prefill_tokens_per_s_q75": 1266.3766126755118,
    "handoff_plus_first_decode_ms": 84.755449,
    "handoff_plus_first_decode_ms_q25": 82.5234705,
    "handoff_plus_first_decode_ms_q75": 86.53717,
    "steady_tpot_ms": 51.931095,
    "steady_tpot_ms_q25": 51.822685750000005,
    "steady_tpot_ms_q75": 51.97059125,
    "steady_tpot_p95_ms": 54.00788675,
    "steady_tpot_p95_ms_q25": 53.69206885,
    "steady_tpot_p95_ms_q75": 55.277183474999994,
    "decode_tokens_per_s": 18.343377692067843,
    "decode_tokens_per_s_q25": 18.31718661497328,
    "decode_tokens_per_s_q75": 18.36791021039229,
    "decode_ms": 817.7338030000001,
    "decode_ms_q25": 816.6430795000006,
    "decode_ms_q75": 818.9047249999999,
    "end_to_end_ms": 7387.4511680000005,
    "end_to_end_ms_q25": 7274.859455,
    "end_to_end_ms_q75": 7391.973685000001,
    "output_tokens_per_s_e2e": 2.1658349593303194,
    "output_tokens_per_s_e2e_q25": 2.164510680218256,
    "output_tokens_per_s_e2e_q75": 2.199882146189351,
    "all_tokens_per_s_e2e": 1108.9074991771236,
    "all_tokens_per_s_e2e_q25": 1108.229468271747,
    "all_tokens_per_s_e2e_q75": 1126.3396588489477,
    "energy_net_psys_j": 250.13882033965706,
    "energy_net_psys_j_q25": 247.2570084883078,
    "energy_net_psys_j_q75": 251.32788909997032,
    "prefill_energy_net_psys_j": 223.29775997963776,
    "prefill_energy_net_psys_j_q25": 220.62440282771084,
    "prefill_energy_net_psys_j_q75": 223.97924204847368,
    "decode_energy_net_psys_j": 26.8410603600193,
    "decode_energy_net_psys_j_q25": 26.63260566059695,
    "decode_energy_net_psys_j_q75": 27.34864705149664,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 6571.898811999999,
    "ttft_ms_q25": 6458.2163755,
    "ttft_ms_q75": 6574.1596835,
    "prefill_tokens_per_s": 1244.08488838431,
    "prefill_tokens_per_s_q25": 1243.657191273974,
    "prefill_tokens_per_s_q75": 1266.3766126755118,
    "handoff_plus_first_decode_ms": 1466.3667899999998,
    "handoff_plus_first_decode_ms_q25": 1463.1771775,
    "handoff_plus_first_decode_ms_q75": 1553.9912279999999,
    "steady_tpot_ms": 1349.0224505,
    "steady_tpot_ms_q25": 1341.510896,
    "steady_tpot_ms_q75": 1349.18147275,
    "steady_tpot_p95_ms": 1350.9115041999999,
    "steady_tpot_p95_ms_q25": 1350.0441313000001,
    "steady_tpot_p95_ms_q75": 1350.93285445,
    "decode_tokens_per_s": 0.738418513627556,
    "decode_tokens_per_s_q25": 0.737738155377583,
    "decode_tokens_per_s_q75": 0.7396567073285333,
    "decode_ms": 20313.683532,
    "decode_ms_q25": 20279.735026000002,
    "decode_ms_q75": 20332.434547499997,
    "end_to_end_ms": 26890.104087,
    "end_to_end_ms_q25": 26740.212273,
    "end_to_end_ms_q75": 26906.594230999995,
    "output_tokens_per_s_e2e": 0.5950144316375178,
    "output_tokens_per_s_e2e_q25": 0.5946499907167409,
    "output_tokens_per_s_e2e_q75": 0.5983685767164327,
    "all_tokens_per_s_e2e": 304.6473889984091,
    "all_tokens_per_s_e2e_q25": 304.4607952469713,
    "all_tokens_per_s_e2e_q75": 306.36471127881356,
    "energy_net_psys_j": 463.83781447519215,
    "energy_net_psys_j_q25": 452.63270597417926,
    "energy_net_psys_j_q75": 464.6260754693129,
    "prefill_energy_net_psys_j": 223.29775997963776,
    "prefill_energy_net_psys_j_q25": 220.62440282771084,
    "prefill_energy_net_psys_j_q75": 223.97924204847368,
    "decode_energy_net_psys_j": 240.5400544955544,
    "decode_energy_net_psys_j_q25": 232.00830314646842,
    "decode_energy_net_psys_j_q75": 240.64683342083927,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.27472750362358983,
    "energy_reduction_vs_gpu": -0.8543215876902224
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 18627.111202999997,
    "ttft_ms_q25": 18620.885055,
    "ttft_ms_q75": 18627.8946935,
    "prefill_tokens_per_s": 438.9301116473289,
    "prefill_tokens_per_s_q25": 438.91165099385455,
    "prefill_tokens_per_s_q75": 439.0769230313261,
    "handoff_plus_first_decode_ms": 1466.3667899999998,
    "handoff_plus_first_decode_ms_q25": 1463.1771775,
    "handoff_plus_first_decode_ms_q75": 1553.9912279999999,
    "steady_tpot_ms": 1349.0224505,
    "steady_tpot_ms_q25": 1341.510896,
    "steady_tpot_ms_q75": 1349.18147275,
    "steady_tpot_p95_ms": 1350.9115041999999,
    "steady_tpot_p95_ms_q25": 1350.0441313000001,
    "steady_tpot_p95_ms_q75": 1350.93285445,
    "decode_tokens_per_s": 0.738418513627556,
    "decode_tokens_per_s_q25": 0.737738155377583,
    "decode_tokens_per_s_q75": 0.7396567073285333,
    "decode_ms": 20313.683532,
    "decode_ms_q25": 20279.735026000002,
    "decode_ms_q75": 20332.434547499997,
    "end_to_end_ms": 38940.794734999996,
    "end_to_end_ms_q25": 38900.620081,
    "end_to_end_ms_q75": 38960.329241,
    "output_tokens_per_s_e2e": 0.4108801607384555,
    "output_tokens_per_s_e2e_q25": 0.4106742508097466,
    "output_tokens_per_s_e2e_q75": 0.411304936319686,
    "all_tokens_per_s_e2e": 210.37064229808922,
    "all_tokens_per_s_e2e_q25": 210.26521641459027,
    "all_tokens_per_s_e2e_q75": 210.58812739567924,
    "energy_net_psys_j": 574.7364770321341,
    "energy_net_psys_j_q25": 560.7616380076554,
    "energy_net_psys_j_q75": 580.8076200823334,
    "prefill_energy_net_psys_j": 334.1964225365797,
    "prefill_energy_net_psys_j_q25": 328.753334861187,
    "prefill_energy_net_psys_j_q75": 340.16078666149417,
    "decode_energy_net_psys_j": 240.5400544955544,
    "decode_energy_net_psys_j_q25": 232.00830314646842,
    "decode_energy_net_psys_j_q75": 240.64683342083927,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.1897098202097082,
    "energy_reduction_vs_gpu": -1.2976700547788393
  },
  {
    "precision": "INT8",
    "model": "qwen25_3b",
    "model_label": "Qwen2.5 3B",
    "family": "Qwen2.5",
    "size_b": 3.0,
    "context_length": 8192,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 18627.111202999997,
    "ttft_ms_q25": 18620.885055,
    "ttft_ms_q75": 18627.8946935,
    "prefill_tokens_per_s": 438.9301116473289,
    "prefill_tokens_per_s_q25": 438.91165099385455,
    "prefill_tokens_per_s_q75": 439.0769230313261,
    "handoff_plus_first_decode_ms": 84.755449,
    "handoff_plus_first_decode_ms_q25": 82.5234705,
    "handoff_plus_first_decode_ms_q75": 86.53717,
    "steady_tpot_ms": 51.931095,
    "steady_tpot_ms_q25": 51.822685750000005,
    "steady_tpot_ms_q75": 51.97059125,
    "steady_tpot_p95_ms": 54.00788675,
    "steady_tpot_p95_ms_q25": 53.69206885,
    "steady_tpot_p95_ms_q75": 55.277183474999994,
    "decode_tokens_per_s": 18.343377692067843,
    "decode_tokens_per_s_q25": 18.31718661497328,
    "decode_tokens_per_s_q75": 18.36791021039229,
    "decode_ms": 817.7338030000001,
    "decode_ms_q25": 816.6430795000006,
    "decode_ms_q75": 818.9047249999999,
    "end_to_end_ms": 19444.23054,
    "end_to_end_ms_q25": 19438.311625,
    "end_to_end_ms_q75": 19445.708695,
    "output_tokens_per_s_e2e": 0.8228661950435813,
    "output_tokens_per_s_e2e_q25": 0.8228036500703717,
    "output_tokens_per_s_e2e_q75": 0.8231168319639813,
    "all_tokens_per_s_e2e": 421.3074918623136,
    "all_tokens_per_s_e2e_q25": 421.2754688360303,
    "all_tokens_per_s_e2e_q75": 421.43581796555844,
    "energy_net_psys_j": 361.037482896599,
    "energy_net_psys_j_q25": 355.38594052178394,
    "energy_net_psys_j_q75": 367.5094337129908,
    "prefill_energy_net_psys_j": 334.1964225365797,
    "prefill_energy_net_psys_j_q25": 328.753334861187,
    "prefill_energy_net_psys_j_q75": 340.16078666149417,
    "decode_energy_net_psys_j": 26.8410603600193,
    "decode_energy_net_psys_j_q25": 26.63260566059695,
    "decode_energy_net_psys_j_q75": 27.34864705149664,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.3799302396051513,
    "energy_reduction_vs_gpu": -0.44334846708861697
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 1024,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 851.337049,
    "ttft_ms_q25": 838.5334135,
    "ttft_ms_q75": 863.5570245,
    "prefill_tokens_per_s": 1184.0198910455265,
    "prefill_tokens_per_s_q25": 1167.4989113211343,
    "prefill_tokens_per_s_q75": 1202.3791164405643,
    "handoff_plus_first_decode_ms": 102.50777099999999,
    "handoff_plus_first_decode_ms_q25": 102.41931799999999,
    "handoff_plus_first_decode_ms_q75": 112.10647599999999,
    "steady_tpot_ms": 95.943353,
    "steady_tpot_ms_q25": 95.5408235,
    "steady_tpot_ms_q75": 96.12107925000001,
    "steady_tpot_p95_ms": 97.09120100000001,
    "steady_tpot_p95_ms_q25": 96.64610217500001,
    "steady_tpot_p95_ms_q75": 97.35144450000001,
    "decode_tokens_per_s": 10.329331942723378,
    "decode_tokens_per_s_q25": 10.311855500704429,
    "decode_tokens_per_s_q75": 10.343555809054653,
    "decode_ms": 1452.1752310000002,
    "decode_ms_q25": 1450.1810250000003,
    "decode_ms_q75": 1454.6405430000002,
    "end_to_end_ms": 2299.523868,
    "end_to_end_ms_q25": 2291.1797505000004,
    "end_to_end_ms_q75": 2313.7380495,
    "output_tokens_per_s_e2e": 6.957962134098623,
    "output_tokens_per_s_e2e_q25": 6.915477698495869,
    "output_tokens_per_s_e2e_q75": 6.983394562420749,
    "all_tokens_per_s_e2e": 445.3095765823119,
    "all_tokens_per_s_e2e_q25": 442.59057270373563,
    "all_tokens_per_s_e2e_q75": 446.9372519949279,
    "energy_net_psys_j": 79.48445138212134,
    "energy_net_psys_j_q25": 79.05736669684897,
    "energy_net_psys_j_q75": 79.7732347152897,
    "prefill_energy_net_psys_j": 32.52628369724302,
    "prefill_energy_net_psys_j_q25": 32.07452214529449,
    "prefill_energy_net_psys_j_q75": 32.70903795495218,
    "decode_energy_net_psys_j": 47.170225835796735,
    "decode_energy_net_psys_j_q25": 46.63711207506516,
    "decode_energy_net_psys_j_q75": 47.51595831228606,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 1024,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 851.337049,
    "ttft_ms_q25": 838.5334135,
    "ttft_ms_q75": 863.5570245,
    "prefill_tokens_per_s": 1184.0198910455265,
    "prefill_tokens_per_s_q25": 1167.4989113211343,
    "prefill_tokens_per_s_q75": 1202.3791164405643,
    "handoff_plus_first_decode_ms": 3113.611299,
    "handoff_plus_first_decode_ms_q25": 3113.051727,
    "handoff_plus_first_decode_ms_q75": 3192.6054135,
    "steady_tpot_ms": 3107.214298,
    "steady_tpot_ms_q25": 3105.5753635,
    "steady_tpot_ms_q75": 3107.3147449999997,
    "steady_tpot_p95_ms": 3109.5429515,
    "steady_tpot_p95_ms_q25": 3109.005176525,
    "steady_tpot_p95_ms_q75": 3116.6707525499996,
    "decode_tokens_per_s": 0.32191372763738574,
    "decode_tokens_per_s_q25": 0.32154929644924435,
    "decode_tokens_per_s_q75": 0.3220333420635425,
    "decode_ms": 46596.335328999994,
    "decode_ms_q25": 46579.03424949999,
    "decode_ms_q75": 46649.205684,
    "end_to_end_ms": 47447.672377999996,
    "end_to_end_ms_q25": 47442.591274,
    "end_to_end_ms_q75": 47487.7390975,
    "output_tokens_per_s_e2e": 0.337213591270258,
    "output_tokens_per_s_e2e_q25": 0.33692931468435716,
    "output_tokens_per_s_e2e_q75": 0.33724971073186794,
    "all_tokens_per_s_e2e": 21.581669841296513,
    "all_tokens_per_s_e2e_q25": 21.56347613979886,
    "all_tokens_per_s_e2e_q75": 21.583981486839548,
    "energy_net_psys_j": 494.3333382928764,
    "energy_net_psys_j_q25": 493.6125967366713,
    "energy_net_psys_j_q75": 495.443663715468,
    "prefill_energy_net_psys_j": 32.52628369724302,
    "prefill_energy_net_psys_j_q25": 32.07452214529449,
    "prefill_energy_net_psys_j_q75": 32.70903795495218,
    "decode_energy_net_psys_j": 461.44154608021506,
    "decode_energy_net_psys_j_q25": 461.3553203336677,
    "decode_energy_net_psys_j_q75": 462.73462576051577,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.04846441885874717,
    "energy_reduction_vs_gpu": -5.219245773193677
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 1024,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 4747.19481,
    "ttft_ms_q25": 4745.7933875,
    "ttft_ms_q75": 4748.121032499999,
    "prefill_tokens_per_s": 212.3359247605851,
    "prefill_tokens_per_s_q25": 212.2945121781214,
    "prefill_tokens_per_s_q75": 212.39864562022592,
    "handoff_plus_first_decode_ms": 3113.611299,
    "handoff_plus_first_decode_ms_q25": 3113.051727,
    "handoff_plus_first_decode_ms_q75": 3192.6054135,
    "steady_tpot_ms": 3107.214298,
    "steady_tpot_ms_q25": 3105.5753635,
    "steady_tpot_ms_q75": 3107.3147449999997,
    "steady_tpot_p95_ms": 3109.5429515,
    "steady_tpot_p95_ms_q25": 3109.005176525,
    "steady_tpot_p95_ms_q75": 3116.6707525499996,
    "decode_tokens_per_s": 0.32191372763738574,
    "decode_tokens_per_s_q25": 0.32154929644924435,
    "decode_tokens_per_s_q75": 0.3220333420635425,
    "decode_ms": 46596.335328999994,
    "decode_ms_q25": 46579.03424949999,
    "decode_ms_q75": 46649.205684,
    "end_to_end_ms": 51343.530138999995,
    "end_to_end_ms_q25": 51327.15528199999,
    "end_to_end_ms_q75": 51394.9990715,
    "output_tokens_per_s_e2e": 0.3116264105074959,
    "output_tokens_per_s_e2e_q25": 0.3113146480229191,
    "output_tokens_per_s_e2e_q75": 0.3117258601332318,
    "all_tokens_per_s_e2e": 19.944090272479738,
    "all_tokens_per_s_e2e_q25": 19.92413747346682,
    "all_tokens_per_s_e2e_q75": 19.950455048526834,
    "energy_net_psys_j": 537.0834476199191,
    "energy_net_psys_j_q25": 536.3672046788387,
    "energy_net_psys_j_q75": 538.6695282595207,
    "prefill_energy_net_psys_j": 75.64190153970398,
    "prefill_energy_net_psys_j_q25": 75.01188434517098,
    "prefill_energy_net_psys_j_q75": 75.93490249900486,
    "decode_energy_net_psys_j": 461.44154608021506,
    "decode_energy_net_psys_j_q25": 461.3553203336677,
    "decode_energy_net_psys_j_q75": 462.73462576051577,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.04478702305382205,
    "energy_reduction_vs_gpu": -5.757088188706135
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 1024,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 4747.19481,
    "ttft_ms_q25": 4745.7933875,
    "ttft_ms_q75": 4748.121032499999,
    "prefill_tokens_per_s": 212.3359247605851,
    "prefill_tokens_per_s_q25": 212.2945121781214,
    "prefill_tokens_per_s_q75": 212.39864562022592,
    "handoff_plus_first_decode_ms": 102.50777099999999,
    "handoff_plus_first_decode_ms_q25": 102.41931799999999,
    "handoff_plus_first_decode_ms_q75": 112.10647599999999,
    "steady_tpot_ms": 95.943353,
    "steady_tpot_ms_q25": 95.5408235,
    "steady_tpot_ms_q75": 96.12107925000001,
    "steady_tpot_p95_ms": 97.09120100000001,
    "steady_tpot_p95_ms_q25": 96.64610217500001,
    "steady_tpot_p95_ms_q75": 97.35144450000001,
    "decode_tokens_per_s": 10.329331942723378,
    "decode_tokens_per_s_q25": 10.311855500704429,
    "decode_tokens_per_s_q75": 10.343555809054653,
    "decode_ms": 1452.1752310000002,
    "decode_ms_q25": 1450.1810250000003,
    "decode_ms_q75": 1454.6405430000002,
    "end_to_end_ms": 6201.222486,
    "end_to_end_ms_q25": 6198.3020575,
    "end_to_end_ms_q75": 6201.360153,
    "output_tokens_per_s_e2e": 2.58013642247507,
    "output_tokens_per_s_e2e_q25": 2.580079146047474,
    "output_tokens_per_s_e2e_q75": 2.581352667799944,
    "all_tokens_per_s_e2e": 165.12873103840448,
    "all_tokens_per_s_e2e_q25": 165.12506534703834,
    "all_tokens_per_s_e2e_q75": 165.20657073919642,
    "energy_net_psys_j": 122.33190177263936,
    "energy_net_psys_j_q25": 122.28772985602635,
    "energy_net_psys_j_q75": 122.57201457407004,
    "prefill_energy_net_psys_j": 75.64190153970398,
    "prefill_energy_net_psys_j_q25": 75.01188434517098,
    "prefill_energy_net_psys_j_q75": 75.93490249900486,
    "decode_energy_net_psys_j": 47.170225835796735,
    "decode_energy_net_psys_j_q25": 46.63711207506516,
    "decode_energy_net_psys_j_q75": 47.51595831228606,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.3708178303860973,
    "energy_reduction_vs_gpu": -0.5390670709234564
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 2048,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1814.903333,
    "ttft_ms_q25": 1789.113373,
    "ttft_ms_q75": 1822.5951925,
    "prefill_tokens_per_s": 1119.6188596122877,
    "prefill_tokens_per_s_q25": 1114.9136135404883,
    "prefill_tokens_per_s_q75": 1135.994145842691,
    "handoff_plus_first_decode_ms": 104.643907,
    "handoff_plus_first_decode_ms_q25": 104.42456150000001,
    "handoff_plus_first_decode_ms_q75": 113.77543600000001,
    "steady_tpot_ms": 94.52278799999999,
    "steady_tpot_ms_q25": 94.46632774999999,
    "steady_tpot_ms_q75": 94.7092935,
    "steady_tpot_p95_ms": 95.9783772,
    "steady_tpot_p95_ms_q25": 95.70110405,
    "steady_tpot_p95_ms_q75": 96.119991575,
    "decode_tokens_per_s": 10.44213945369756,
    "decode_tokens_per_s_q25": 10.393711447950494,
    "decode_tokens_per_s_q75": 10.470220632596908,
    "decode_ms": 1436.487232,
    "decode_ms_q25": 1432.6448719999999,
    "decode_ms_q75": 1443.2116695,
    "end_to_end_ms": 3243.705845,
    "end_to_end_ms_q25": 3228.4826825,
    "end_to_end_ms_q75": 3255.2400645,
    "output_tokens_per_s_e2e": 4.932629765014959,
    "output_tokens_per_s_e2e_q25": 4.915213797618445,
    "output_tokens_per_s_e2e_q75": 4.955998628935779,
    "all_tokens_per_s_e2e": 631.3766099219148,
    "all_tokens_per_s_e2e_q25": 629.1473660951609,
    "all_tokens_per_s_e2e_q75": 634.3678245037797,
    "energy_net_psys_j": 118.64620400500446,
    "energy_net_psys_j_q25": 117.51494607654138,
    "energy_net_psys_j_q75": 119.99973734789285,
    "prefill_energy_net_psys_j": 71.53034383558204,
    "prefill_energy_net_psys_j_q25": 70.20090348347841,
    "prefill_energy_net_psys_j_q75": 73.08926092298307,
    "decode_energy_net_psys_j": 47.11586016942242,
    "decode_energy_net_psys_j_q25": 46.91047642490979,
    "decode_energy_net_psys_j_q75": 47.314042593062965,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 2048,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 1814.903333,
    "ttft_ms_q25": 1789.113373,
    "ttft_ms_q75": 1822.5951925,
    "prefill_tokens_per_s": 1119.6188596122877,
    "prefill_tokens_per_s_q25": 1114.9136135404883,
    "prefill_tokens_per_s_q75": 1135.994145842691,
    "handoff_plus_first_decode_ms": 3152.366479,
    "handoff_plus_first_decode_ms_q25": 3137.942583,
    "handoff_plus_first_decode_ms_q75": 3207.458847,
    "steady_tpot_ms": 3106.8281524999998,
    "steady_tpot_ms_q25": 3106.2449895,
    "steady_tpot_ms_q75": 3108.384433,
    "steady_tpot_p95_ms": 3111.3768621000004,
    "steady_tpot_p95_ms_q25": 3111.1016603000003,
    "steady_tpot_p95_ms_q75": 3117.6600318250003,
    "decode_tokens_per_s": 0.32192174323489836,
    "decode_tokens_per_s_q25": 0.32110542449324053,
    "decode_tokens_per_s_q75": 0.3220267201852633,
    "decode_ms": 46595.175117000006,
    "decode_ms_q25": 46579.990584,
    "decode_ms_q75": 46713.931934,
    "end_to_end_ms": 48425.462169000006,
    "end_to_end_ms_q25": 48402.585776500004,
    "end_to_end_ms_q75": 48510.737166499996,
    "output_tokens_per_s_e2e": 0.3304046937985146,
    "output_tokens_per_s_e2e_q25": 0.32982490837615175,
    "output_tokens_per_s_e2e_q75": 0.3305609259783978,
    "all_tokens_per_s_e2e": 42.29180080620987,
    "all_tokens_per_s_e2e_q25": 42.217588272147424,
    "all_tokens_per_s_e2e_q75": 42.31179852523492,
    "energy_net_psys_j": 554.6957757140775,
    "energy_net_psys_j_q25": 541.3838384201301,
    "energy_net_psys_j_q75": 558.1248563589727,
    "prefill_energy_net_psys_j": 71.53034383558204,
    "prefill_energy_net_psys_j_q25": 70.20090348347841,
    "prefill_energy_net_psys_j_q75": 73.08926092298307,
    "decode_energy_net_psys_j": 480.0475977036934,
    "decode_energy_net_psys_j_q25": 468.294577497147,
    "decode_energy_net_psys_j_q75": 486.3650357880932,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.06698347728060482,
    "energy_reduction_vs_gpu": -3.6752087887336087
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 2048,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 9483.739544,
    "ttft_ms_q25": 9479.5676275,
    "ttft_ms_q75": 9492.238798,
    "prefill_tokens_per_s": 214.26147255230862,
    "prefill_tokens_per_s_q25": 214.06979663537786,
    "prefill_tokens_per_s_q75": 214.35580961219705,
    "handoff_plus_first_decode_ms": 3152.366479,
    "handoff_plus_first_decode_ms_q25": 3137.942583,
    "handoff_plus_first_decode_ms_q75": 3207.458847,
    "steady_tpot_ms": 3106.8281524999998,
    "steady_tpot_ms_q25": 3106.2449895,
    "steady_tpot_ms_q75": 3108.384433,
    "steady_tpot_p95_ms": 3111.3768621000004,
    "steady_tpot_p95_ms_q25": 3111.1016603000003,
    "steady_tpot_p95_ms_q75": 3117.6600318250003,
    "decode_tokens_per_s": 0.32192174323489836,
    "decode_tokens_per_s_q25": 0.32110542449324053,
    "decode_tokens_per_s_q75": 0.3220267201852633,
    "decode_ms": 46595.175117000006,
    "decode_ms_q25": 46579.990584,
    "decode_ms_q75": 46713.931934,
    "end_to_end_ms": 56070.57082800001,
    "end_to_end_ms_q25": 56068.05746550001,
    "end_to_end_ms_q75": 56193.4995615,
    "output_tokens_per_s_e2e": 0.285354683637536,
    "output_tokens_per_s_e2e_q25": 0.28473180518235236,
    "output_tokens_per_s_e2e_q75": 0.2853674758037859,
    "all_tokens_per_s_e2e": 36.52539950560461,
    "all_tokens_per_s_e2e_q25": 36.4456710633411,
    "all_tokens_per_s_e2e_q75": 36.527036902884596,
    "energy_net_psys_j": 641.1556664126668,
    "energy_net_psys_j_q25": 622.671539129447,
    "energy_net_psys_j_q75": 644.4096714536097,
    "prefill_energy_net_psys_j": 154.98120262205958,
    "prefill_energy_net_psys_j_q25": 151.3135285888431,
    "prefill_energy_net_psys_j_q75": 158.04463566551652,
    "decode_energy_net_psys_j": 480.0475977036934,
    "decode_energy_net_psys_j_q25": 468.294577497147,
    "decode_energy_net_psys_j_q75": 486.3650357880932,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.057850415950825095,
    "energy_reduction_vs_gpu": -4.403929032450319
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 2048,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 9483.739544,
    "ttft_ms_q25": 9479.5676275,
    "ttft_ms_q75": 9492.238798,
    "prefill_tokens_per_s": 214.26147255230862,
    "prefill_tokens_per_s_q25": 214.06979663537786,
    "prefill_tokens_per_s_q75": 214.35580961219705,
    "handoff_plus_first_decode_ms": 104.643907,
    "handoff_plus_first_decode_ms_q25": 104.42456150000001,
    "handoff_plus_first_decode_ms_q75": 113.77543600000001,
    "steady_tpot_ms": 94.52278799999999,
    "steady_tpot_ms_q25": 94.46632774999999,
    "steady_tpot_ms_q75": 94.7092935,
    "steady_tpot_p95_ms": 95.9783772,
    "steady_tpot_p95_ms_q25": 95.70110405,
    "steady_tpot_p95_ms_q75": 96.119991575,
    "decode_tokens_per_s": 10.44213945369756,
    "decode_tokens_per_s_q25": 10.393711447950494,
    "decode_tokens_per_s_q75": 10.470220632596908,
    "decode_ms": 1436.487232,
    "decode_ms_q25": 1432.6448719999999,
    "decode_ms_q75": 1443.2116695,
    "end_to_end_ms": 10929.540564,
    "end_to_end_ms_q25": 10920.7117535,
    "end_to_end_ms_q75": 10931.6081075,
    "output_tokens_per_s_e2e": 1.4639224683150183,
    "output_tokens_per_s_e2e_q25": 1.4636456425264657,
    "output_tokens_per_s_e2e_q75": 1.4651069287905347,
    "all_tokens_per_s_e2e": 187.38207594432234,
    "all_tokens_per_s_e2e_q25": 187.3466422433876,
    "all_tokens_per_s_e2e_q75": 187.53368688518844,
    "energy_net_psys_j": 202.49342763876308,
    "energy_net_psys_j_q25": 198.62757118190606,
    "energy_net_psys_j_q75": 205.15329451406686,
    "prefill_energy_net_psys_j": 154.98120262205958,
    "prefill_energy_net_psys_j_q25": 151.3135285888431,
    "prefill_energy_net_psys_j_q75": 158.04463566551652,
    "decode_energy_net_psys_j": 47.11586016942242,
    "decode_energy_net_psys_j_q25": 46.91047642490979,
    "decode_energy_net_psys_j_q75": 47.314042593062965,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.29678336669376576,
    "energy_reduction_vs_gpu": -0.7066995892276668
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 4096,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 4301.698888,
    "ttft_ms_q25": 4259.214180999999,
    "ttft_ms_q75": 4308.775047499999,
    "prefill_tokens_per_s": 948.4624810401189,
    "prefill_tokens_per_s_q25": 946.9074061606327,
    "prefill_tokens_per_s_q75": 958.0185009024171,
    "handoff_plus_first_decode_ms": 113.064014,
    "handoff_plus_first_decode_ms_q25": 112.7966735,
    "handoff_plus_first_decode_ms_q75": 120.5083855,
    "steady_tpot_ms": 98.9142945,
    "steady_tpot_ms_q25": 98.7371235,
    "steady_tpot_ms_q75": 98.9985035,
    "steady_tpot_p95_ms": 99.9681923,
    "steady_tpot_p95_ms_q25": 99.924901175,
    "steady_tpot_p95_ms_q75": 101.56567347500001,
    "decode_tokens_per_s": 9.988896282958486,
    "decode_tokens_per_s_q25": 9.916204961946324,
    "decode_tokens_per_s_q75": 10.005481032816757,
    "decode_ms": 1501.6674090000006,
    "decode_ms_q25": 1499.1824145000005,
    "decode_ms_q75": 1512.7567610000006,
    "end_to_end_ms": 5803.3662970000005,
    "end_to_end_ms_q25": 5771.970942,
    "end_to_end_ms_q75": 5807.957462,
    "output_tokens_per_s_e2e": 2.757020525874967,
    "output_tokens_per_s_e2e_q25": 2.7548428346494895,
    "output_tokens_per_s_e2e_q75": 2.772098741785606,
    "all_tokens_per_s_e2e": 705.7972546239915,
    "all_tokens_per_s_e2e_q25": 705.2397656702693,
    "all_tokens_per_s_e2e_q75": 709.6572778971151,
    "energy_net_psys_j": 211.7978605549249,
    "energy_net_psys_j_q25": 210.96046435290813,
    "energy_net_psys_j_q75": 213.74874481832757,
    "prefill_energy_net_psys_j": 164.6064481677537,
    "prefill_energy_net_psys_j_q25": 162.7087105495131,
    "prefill_energy_net_psys_j_q75": 165.48876207032595,
    "decode_energy_net_psys_j": 49.31209521961884,
    "decode_energy_net_psys_j_q25": 48.25175380339502,
    "decode_energy_net_psys_j_q75": 49.320324164225454,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 4096,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 4301.698888,
    "ttft_ms_q25": 4259.214180999999,
    "ttft_ms_q75": 4308.775047499999,
    "prefill_tokens_per_s": 948.4624810401189,
    "prefill_tokens_per_s_q25": 946.9074061606327,
    "prefill_tokens_per_s_q75": 958.0185009024171,
    "handoff_plus_first_decode_ms": 3243.338733,
    "handoff_plus_first_decode_ms_q25": 3224.3841645,
    "handoff_plus_first_decode_ms_q75": 3268.4665535,
    "steady_tpot_ms": 3112.18156,
    "steady_tpot_ms_q25": 3110.20444125,
    "steady_tpot_ms_q75": 3112.53361825,
    "steady_tpot_p95_ms": 3116.2414857000003,
    "steady_tpot_p95_ms_q25": 3115.6718375,
    "steady_tpot_p95_ms_q75": 3118.8060249500004,
    "decode_tokens_per_s": 0.32070852688149143,
    "decode_tokens_per_s_q25": 0.32045757980661094,
    "decode_tokens_per_s_q75": 0.3213179714875056,
    "decode_ms": 46771.44117699999,
    "decode_ms_q25": 46682.897607499996,
    "decode_ms_q75": 46808.0961215,
    "end_to_end_ms": 51061.48054000001,
    "end_to_end_ms_q25": 50985.842892500004,
    "end_to_end_ms_q75": 51067.3103025,
    "output_tokens_per_s_e2e": 0.31334774923860825,
    "output_tokens_per_s_e2e_q25": 0.3133119820436474,
    "output_tokens_per_s_e2e_q75": 0.31381329217813825,
    "all_tokens_per_s_e2e": 80.21702380508371,
    "all_tokens_per_s_e2e_q25": 80.20786740317374,
    "all_tokens_per_s_e2e_q75": 80.33620279760339,
    "energy_net_psys_j": 657.1732641951776,
    "energy_net_psys_j_q25": 639.5415244349463,
    "energy_net_psys_j_q75": 665.8276238963842,
    "prefill_energy_net_psys_j": 164.6064481677537,
    "prefill_energy_net_psys_j_q25": 162.7087105495131,
    "prefill_energy_net_psys_j_q75": 165.48876207032595,
    "decode_energy_net_psys_j": 496.36229126390504,
    "decode_energy_net_psys_j_q25": 476.83281388543315,
    "decode_energy_net_psys_j_q75": 502.2365994442989,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.11365448544825918,
    "energy_reduction_vs_gpu": -2.1028324010135826
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 4096,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 19791.634002,
    "ttft_ms_q25": 19780.841848,
    "ttft_ms_q75": 19794.6949645,
    "prefill_tokens_per_s": 206.14770865243895,
    "prefill_tokens_per_s_q25": 206.11583582766158,
    "prefill_tokens_per_s_q75": 206.26024139023977,
    "handoff_plus_first_decode_ms": 3243.338733,
    "handoff_plus_first_decode_ms_q25": 3224.3841645,
    "handoff_plus_first_decode_ms_q75": 3268.4665535,
    "steady_tpot_ms": 3112.18156,
    "steady_tpot_ms_q25": 3110.20444125,
    "steady_tpot_ms_q75": 3112.53361825,
    "steady_tpot_p95_ms": 3116.2414857000003,
    "steady_tpot_p95_ms_q25": 3115.6718375,
    "steady_tpot_p95_ms_q75": 3118.8060249500004,
    "decode_tokens_per_s": 0.32070852688149143,
    "decode_tokens_per_s_q25": 0.32045757980661094,
    "decode_tokens_per_s_q75": 0.3213179714875056,
    "decode_ms": 46771.44117699999,
    "decode_ms_q25": 46682.897607499996,
    "decode_ms_q75": 46808.0961215,
    "end_to_end_ms": 66563.07517899999,
    "end_to_end_ms_q25": 66477.592572,
    "end_to_end_ms_q75": 66588.9379695,
    "output_tokens_per_s_e2e": 0.24037350974204758,
    "output_tokens_per_s_e2e_q25": 0.24028018619152602,
    "output_tokens_per_s_e2e_q75": 0.24068300064808024,
    "all_tokens_per_s_e2e": 61.53561849396418,
    "all_tokens_per_s_e2e_q25": 61.51172766503066,
    "all_tokens_per_s_e2e_q75": 61.61484816590854,
    "energy_net_psys_j": 821.4738069699072,
    "energy_net_psys_j_q25": 803.8553449388226,
    "energy_net_psys_j_q75": 833.5423562237002,
    "prefill_energy_net_psys_j": 328.9335464007768,
    "prefill_energy_net_psys_j_q25": 327.02253105338946,
    "prefill_energy_net_psys_j_q75": 333.2167721267886,
    "decode_energy_net_psys_j": 496.36229126390504,
    "decode_energy_net_psys_j_q25": 476.83281388543315,
    "decode_energy_net_psys_j_q75": 502.2365994442989,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.08718597032053751,
    "energy_reduction_vs_gpu": -2.8785746221306936
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 4096,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 19791.634002,
    "ttft_ms_q25": 19780.841848,
    "ttft_ms_q75": 19794.6949645,
    "prefill_tokens_per_s": 206.14770865243895,
    "prefill_tokens_per_s_q25": 206.11583582766158,
    "prefill_tokens_per_s_q75": 206.26024139023977,
    "handoff_plus_first_decode_ms": 113.064014,
    "handoff_plus_first_decode_ms_q25": 112.7966735,
    "handoff_plus_first_decode_ms_q75": 120.5083855,
    "steady_tpot_ms": 98.9142945,
    "steady_tpot_ms_q25": 98.7371235,
    "steady_tpot_ms_q75": 98.9985035,
    "steady_tpot_p95_ms": 99.9681923,
    "steady_tpot_p95_ms_q25": 99.924901175,
    "steady_tpot_p95_ms_q75": 101.56567347500001,
    "decode_tokens_per_s": 9.988896282958486,
    "decode_tokens_per_s_q25": 9.916204961946324,
    "decode_tokens_per_s_q75": 10.005481032816757,
    "decode_ms": 1501.6674090000006,
    "decode_ms_q25": 1499.1824145000005,
    "decode_ms_q75": 1512.7567610000006,
    "end_to_end_ms": 21293.895807,
    "end_to_end_ms_q25": 21293.598609,
    "end_to_end_ms_q75": 21294.174577,
    "output_tokens_per_s_e2e": 0.7513890433680189,
    "output_tokens_per_s_e2e_q25": 0.7513792067819274,
    "output_tokens_per_s_e2e_q75": 0.7513995307657503,
    "all_tokens_per_s_e2e": 192.35559510221285,
    "all_tokens_per_s_e2e_q25": 192.3530769361734,
    "all_tokens_per_s_e2e_q75": 192.35827987603207,
    "energy_net_psys_j": 376.124958787948,
    "energy_net_psys_j_q25": 375.2742848567845,
    "energy_net_psys_j_q75": 381.47675487479023,
    "prefill_energy_net_psys_j": 328.9335464007768,
    "prefill_energy_net_psys_j_q25": 327.02253105338946,
    "prefill_energy_net_psys_j_q75": 333.2167721267886,
    "decode_energy_net_psys_j": 49.31209521961884,
    "decode_energy_net_psys_j_q25": 48.25175380339502,
    "decode_energy_net_psys_j_q75": 49.320324164225454,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.27253661563856457,
    "energy_reduction_vs_gpu": -0.7758676022622459
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 8192,
    "route": "gpu_full",
    "route_label": "Whole iGPU",
    "mode": "measured",
    "prefill_device": "iGPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 10684.807526,
    "ttft_ms_q25": 10595.7277625,
    "ttft_ms_q75": 10686.7457425,
    "prefill_tokens_per_s": 765.1986224463881,
    "prefill_tokens_per_s_q25": 765.0598663000288,
    "prefill_tokens_per_s_q75": 771.6862969477645,
    "handoff_plus_first_decode_ms": 144.399666,
    "handoff_plus_first_decode_ms_q25": 143.42664050000002,
    "handoff_plus_first_decode_ms_q75": 148.60709200000002,
    "steady_tpot_ms": 106.494876,
    "steady_tpot_ms_q25": 106.358266,
    "steady_tpot_ms_q75": 106.84022300000001,
    "steady_tpot_p95_ms": 108.0899779,
    "steady_tpot_p95_ms_q25": 108.086983825,
    "steady_tpot_p95_ms_q75": 108.29293767499999,
    "decode_tokens_per_s": 9.118984476622629,
    "decode_tokens_per_s_q25": 9.111874753983312,
    "decode_tokens_per_s_q75": 9.140002438818653,
    "decode_ms": 1644.9200059999985,
    "decode_ms_q25": 1641.1460959999995,
    "decode_ms_q75": 1646.2044899999992,
    "end_to_end_ms": 12322.179712000001,
    "end_to_end_ms_q25": 12236.8738585,
    "end_to_end_ms_q75": 12329.1763225,
    "output_tokens_per_s_e2e": 1.2984715670408815,
    "output_tokens_per_s_e2e_q25": 1.297735123100832,
    "output_tokens_per_s_e2e_q75": 1.3075870347094432,
    "all_tokens_per_s_e2e": 664.8174423249313,
    "all_tokens_per_s_e2e_q25": 664.440383027626,
    "all_tokens_per_s_e2e_q75": 669.4845617712349,
    "energy_net_psys_j": 438.76027407536543,
    "energy_net_psys_j_q25": 437.66146994311407,
    "energy_net_psys_j_q75": 443.1595887379537,
    "prefill_energy_net_psys_j": 386.72118574907455,
    "prefill_energy_net_psys_j_q25": 385.18148028533506,
    "prefill_energy_net_psys_j_q75": 391.27069725958825,
    "decode_energy_net_psys_j": 52.03908832629088,
    "decode_energy_net_psys_j_q25": 51.88889147836545,
    "decode_energy_net_psys_j_q75": 52.47998965777896,
    "fastest_e2e": true,
    "lowest_energy": true,
    "e2e_speedup_vs_gpu": 1.0,
    "energy_reduction_vs_gpu": 0.0
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 8192,
    "route": "gpu_npu_oracle",
    "route_label": "iGPU prefill + NPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "iGPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 10684.807526,
    "ttft_ms_q25": 10595.7277625,
    "ttft_ms_q75": 10686.7457425,
    "prefill_tokens_per_s": 765.1986224463881,
    "prefill_tokens_per_s_q25": 765.0598663000288,
    "prefill_tokens_per_s_q75": 771.6862969477645,
    "handoff_plus_first_decode_ms": 3322.44694,
    "handoff_plus_first_decode_ms_q25": 3316.6843965,
    "handoff_plus_first_decode_ms_q75": 3356.4823385,
    "steady_tpot_ms": 3127.600312,
    "steady_tpot_ms_q25": 3123.33519125,
    "steady_tpot_ms_q75": 3128.7201692500003,
    "steady_tpot_p95_ms": 3131.83723055,
    "steady_tpot_p95_ms_q25": 3131.3378924,
    "steady_tpot_p95_ms_q75": 3135.3172826750006,
    "decode_tokens_per_s": 0.3186405682687098,
    "decode_tokens_per_s_q25": 0.318304801402975,
    "decode_tokens_per_s_q75": 0.3192465858809851,
    "decode_ms": 47074.985089,
    "decode_ms_q25": 46985.793162,
    "decode_ms_q75": 47124.6950245,
    "end_to_end_ms": 57681.05295899999,
    "end_to_end_ms_q25": 57631.230859999996,
    "end_to_end_ms_q75": 57722.3610035,
    "output_tokens_per_s_e2e": 0.27738744664340453,
    "output_tokens_per_s_e2e_q25": 0.27718908090281824,
    "output_tokens_per_s_e2e_q75": 0.27762745509622627,
    "all_tokens_per_s_e2e": 142.02237268142312,
    "all_tokens_per_s_e2e_q25": 141.92080942224294,
    "all_tokens_per_s_e2e_q75": 142.14525700926785,
    "energy_net_psys_j": 897.8886979600086,
    "energy_net_psys_j_q25": 890.2210642062861,
    "energy_net_psys_j_q75": 905.4720977666332,
    "prefill_energy_net_psys_j": 386.72118574907455,
    "prefill_energy_net_psys_j_q25": 385.18148028533506,
    "prefill_energy_net_psys_j_q75": 391.27069725958825,
    "decode_energy_net_psys_j": 511.1675122109341,
    "decode_energy_net_psys_j_q25": 498.95036694669784,
    "decode_energy_net_psys_j_q75": 520.2906174812981,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.2136261229620526,
    "energy_reduction_vs_gpu": -1.046422046417491
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 8192,
    "route": "npu_full",
    "route_label": "Whole NPU",
    "mode": "measured",
    "prefill_device": "NPU",
    "decode_device": "NPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 41954.709656,
    "ttft_ms_q25": 41946.1006585,
    "ttft_ms_q75": 41962.1950545,
    "prefill_tokens_per_s": 194.87681042337374,
    "prefill_tokens_per_s_q25": 194.84205365208828,
    "prefill_tokens_per_s_q75": 194.91681505663374,
    "handoff_plus_first_decode_ms": 3322.44694,
    "handoff_plus_first_decode_ms_q25": 3316.6843965,
    "handoff_plus_first_decode_ms_q75": 3356.4823385,
    "steady_tpot_ms": 3127.600312,
    "steady_tpot_ms_q25": 3123.33519125,
    "steady_tpot_ms_q75": 3128.7201692500003,
    "steady_tpot_p95_ms": 3131.83723055,
    "steady_tpot_p95_ms_q25": 3131.3378924,
    "steady_tpot_p95_ms_q75": 3135.3172826750006,
    "decode_tokens_per_s": 0.3186405682687098,
    "decode_tokens_per_s_q25": 0.318304801402975,
    "decode_tokens_per_s_q75": 0.3192465858809851,
    "decode_ms": 47074.985089,
    "decode_ms_q25": 46985.793162,
    "decode_ms_q75": 47124.6950245,
    "end_to_end_ms": 89012.47675,
    "end_to_end_ms_q25": 88931.8938205,
    "end_to_end_ms_q75": 89078.2810815,
    "output_tokens_per_s_e2e": 0.1797500820580189,
    "output_tokens_per_s_e2e_q25": 0.17961739424219691,
    "output_tokens_per_s_e2e_q75": 0.17991310484344375,
    "all_tokens_per_s_e2e": 92.03204201370568,
    "all_tokens_per_s_e2e_q25": 91.96410585200482,
    "all_tokens_per_s_e2e_q75": 92.1155096798432,
    "energy_net_psys_j": 1275.755816067187,
    "energy_net_psys_j_q25": 1259.7978716065495,
    "energy_net_psys_j_q75": 1276.8386873869003,
    "prefill_energy_net_psys_j": 757.1067054634507,
    "prefill_energy_net_psys_j_q25": 752.8072707092012,
    "prefill_energy_net_psys_j_q75": 760.8475046598518,
    "decode_energy_net_psys_j": 511.1675122109341,
    "decode_energy_net_psys_j_q25": 498.95036694669784,
    "decode_energy_net_psys_j_q75": 520.2906174812981,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.1384320508978535,
    "energy_reduction_vs_gpu": -1.907637476423063
  },
  {
    "precision": "INT8",
    "model": "qwen25_7b",
    "model_label": "Qwen2.5 7B",
    "family": "Qwen2.5",
    "size_b": 7.0,
    "context_length": 8192,
    "route": "npu_gpu_oracle",
    "route_label": "NPU prefill + iGPU decode \u00b7 oracle",
    "mode": "overhead_free_oracle",
    "prefill_device": "NPU",
    "decode_device": "iGPU",
    "repeats": 3,
    "internal_overhead_ms": 0.0,
    "ttft_ms": 41954.709656,
    "ttft_ms_q25": 41946.1006585,
    "ttft_ms_q75": 41962.1950545,
    "prefill_tokens_per_s": 194.87681042337374,
    "prefill_tokens_per_s_q25": 194.84205365208828,
    "prefill_tokens_per_s_q75": 194.91681505663374,
    "handoff_plus_first_decode_ms": 144.399666,
    "handoff_plus_first_decode_ms_q25": 143.42664050000002,
    "handoff_plus_first_decode_ms_q75": 148.60709200000002,
    "steady_tpot_ms": 106.494876,
    "steady_tpot_ms_q25": 106.358266,
    "steady_tpot_ms_q75": 106.84022300000001,
    "steady_tpot_p95_ms": 108.0899779,
    "steady_tpot_p95_ms_q25": 108.086983825,
    "steady_tpot_p95_ms_q75": 108.29293767499999,
    "decode_tokens_per_s": 9.118984476622629,
    "decode_tokens_per_s_q25": 9.111874753983312,
    "decode_tokens_per_s_q75": 9.140002438818653,
    "decode_ms": 1644.9200059999985,
    "decode_ms_q25": 1641.1460959999995,
    "decode_ms_q75": 1646.2044899999992,
    "end_to_end_ms": 43592.081842,
    "end_to_end_ms_q25": 43588.5312385,
    "end_to_end_ms_q75": 43603.3411505,
    "output_tokens_per_s_e2e": 0.3670391347215805,
    "output_tokens_per_s_e2e_q25": 0.36694438189097073,
    "output_tokens_per_s_e2e_q75": 0.3670690351693355,
    "all_tokens_per_s_e2e": 187.92403697744922,
    "all_tokens_per_s_e2e_q25": 187.875523528177,
    "all_tokens_per_s_e2e_q75": 187.93934600669976,
    "energy_net_psys_j": 808.8454000938907,
    "energy_net_psys_j_q25": 805.1370635190547,
    "energy_net_psys_j_q75": 812.7363961382173,
    "prefill_energy_net_psys_j": 757.1067054634507,
    "prefill_energy_net_psys_j_q25": 752.8072707092012,
    "prefill_energy_net_psys_j_q75": 760.8475046598518,
    "decode_energy_net_psys_j": 52.03908832629088,
    "decode_energy_net_psys_j_q25": 51.88889147836545,
    "decode_energy_net_psys_j_q75": 52.47998965777896,
    "fastest_e2e": false,
    "lowest_energy": false,
    "e2e_speedup_vs_gpu": 0.2826701362110184,
    "energy_reduction_vs_gpu": -0.8434791112263644
  }
]