{
  "rows": [
    {
      "engine": "llama.cpp IQ4_XS + MTP2",
      "kind": "ladder",
      "target_tokens": 0,
      "input_tokens": 11,
      "computed_prompt_tokens": 11,
      "prefill_tok_s": null,
      "decode_tok_s": 55.88150927286242,
      "decode_min": 55.858028238441875,
      "decode_max": 56.01341271725382,
      "draft_accept_rate": 0.956
    },
    {
      "engine": "llama.cpp IQ4_XS + MTP2",
      "kind": "ladder",
      "target_tokens": 8000,
      "input_tokens": 7887,
      "computed_prompt_tokens": 7887,
      "prefill_tok_s": 963.4923500881888,
      "decode_tok_s": 56.2957402889848,
      "decode_min": 56.18433455898418,
      "decode_max": 56.399703740655305,
      "draft_accept_rate": 0.997
    },
    {
      "engine": "llama.cpp IQ4_XS + MTP2",
      "kind": "ladder",
      "target_tokens": 16000,
      "input_tokens": 16050,
      "computed_prompt_tokens": 8679,
      "prefill_tok_s": 762.8448249341883,
      "decode_tok_s": 53.81374336209631,
      "decode_min": 53.710861368449514,
      "decode_max": 54.81682401308318,
      "draft_accept_rate": 0.997
    },
    {
      "engine": "llama.cpp IQ4_XS + MTP2",
      "kind": "ladder",
      "target_tokens": 32000,
      "input_tokens": 32460,
      "computed_prompt_tokens": 16926,
      "prefill_tok_s": 573.9593561767077,
      "decode_tok_s": 47.88810577375243,
      "decode_min": 47.81638050389754,
      "decode_max": 48.69840473483036,
      "draft_accept_rate": 0.987
    },
    {
      "engine": "llama.cpp IQ4_XS + MTP2",
      "kind": "pp0",
      "target_tokens": 512,
      "input_tokens": 512,
      "prefill_tok_s": 876.7603704312565
    },
    {
      "engine": "llama.cpp IQ4_XS + MTP2",
      "kind": "pp0",
      "target_tokens": 2048,
      "input_tokens": 1978,
      "prefill_tok_s": 972.094785630493
    },
    {
      "engine": "llama.cpp IQ4_XS + MTP2",
      "kind": "pp0",
      "target_tokens": 8192,
      "input_tokens": 8077,
      "prefill_tok_s": 828.1915289039048
    },
    {
      "engine": "llama.cpp IQ4_XS + MTP2",
      "kind": "real",
      "case": "design",
      "input_tokens": 73,
      "output_tokens": 1200,
      "decode_tok_s": 47.14622596623811,
      "decode_min": 45.77665847039224,
      "decode_max": 49.561146177769174,
      "draft_accept_rate": 0.645
    },
    {
      "engine": "llama.cpp IQ4_XS + MTP2",
      "kind": "real",
      "case": "review",
      "input_tokens": 58,
      "output_tokens": 1200,
      "decode_tok_s": 41.01724126724461,
      "decode_min": 40.6952265245988,
      "decode_max": 44.813257847225664,
      "draft_accept_rate": 0.521
    },
    {
      "engine": "llama.cpp IQ4_XS + MTP2",
      "kind": "real",
      "case": "qa",
      "input_tokens": 61,
      "output_tokens": 1200,
      "decode_tok_s": 45.565598862129335,
      "decode_min": 44.9065638227703,
      "decode_max": 46.25502676214002,
      "draft_accept_rate": 0.622
    },
    {
      "engine": "GGZ14 MXFP4 + DFlash2-7",
      "kind": "ladder",
      "case": "ladder-0",
      "target_tokens": 0,
      "input_tokens": 11,
      "output_tokens": 1000,
      "computed_prompt_tokens": 11.0,
      "prefill_tok_s": null,
      "decode_tok_s": 47.82176754719164,
      "api_tail_tok_s": 47.82225504634838,
      "api_ttft_s": 0.043322253972291946,
      "api_e2e_s": 20.933246277971193,
      "draft_accept_rate": 0.09723889555822329,
      "decode_min": 47.8139548482826,
      "decode_max": 47.82342905471188
    },
    {
      "engine": "GGZ14 MXFP4 + DFlash2-7",
      "kind": "ladder",
      "case": "ladder-8000",
      "target_tokens": 8000,
      "input_tokens": 7887,
      "output_tokens": 1000,
      "computed_prompt_tokens": 7887.0,
      "prefill_tok_s": 3362.740783458321,
      "decode_tok_s": 216.71825269167465,
      "api_tail_tok_s": 216.8892398744369,
      "api_ttft_s": 2.3533540298230946,
      "api_e2e_s": 6.960947191808373,
      "draft_accept_rate": 0.9965986394557823,
      "decode_min": 216.64093420022724,
      "decode_max": 216.71888967812117
    },
    {
      "engine": "GGZ14 MXFP4 + DFlash2-7",
      "kind": "ladder",
      "case": "ladder-16000",
      "target_tokens": 16000,
      "input_tokens": 16050,
      "output_tokens": 1000,
      "computed_prompt_tokens": 9890.0,
      "prefill_tok_s": 3167.6549734837363,
      "decode_tok_s": 213.80803480499495,
      "api_tail_tok_s": 214.27567499889523,
      "api_ttft_s": 3.136147879064083,
      "api_e2e_s": 7.800507094012573,
      "draft_accept_rate": 0.9965986394557823,
      "decode_min": 213.8009294372368,
      "decode_max": 213.87242989304102
    },
    {
      "engine": "GGZ14 MXFP4 + DFlash2-7",
      "kind": "ladder",
      "case": "ladder-32000",
      "target_tokens": 32000,
      "input_tokens": 32460,
      "output_tokens": 1000,
      "computed_prompt_tokens": 17500.0,
      "prefill_tok_s": 2894.887315443021,
      "decode_tok_s": 192.48322372036446,
      "api_tail_tok_s": 193.06936366006403,
      "api_ttft_s": 6.070800324901938,
      "api_e2e_s": 11.243048965930939,
      "draft_accept_rate": 0.9086134453781513,
      "decode_min": 174.60583404262042,
      "decode_max": 192.48859705302118
    },
    {
      "engine": "GGZ14 MXFP4 + DFlash2-7",
      "kind": "pp0",
      "case": "pp512",
      "target_tokens": 512,
      "input_tokens": 512,
      "output_tokens": 64,
      "computed_prompt_tokens": 512.0,
      "prefill_tok_s": 2952.7207294276886,
      "decode_tok_s": 198.25578527751608,
      "api_tail_tok_s": 198.41704214014746,
      "api_ttft_s": 0.17672426998615265,
      "api_e2e_s": 0.49432825087569654,
      "draft_accept_rate": 0.9365079365079365,
      "decode_min": 198.1609324741391,
      "decode_max": 198.31710237818328
    },
    {
      "engine": "GGZ14 MXFP4 + DFlash2-7",
      "kind": "pp0",
      "case": "pp2048",
      "target_tokens": 2048,
      "input_tokens": 1978,
      "output_tokens": 64,
      "computed_prompt_tokens": 1978.0,
      "prefill_tok_s": 3121.0870431272087,
      "decode_tok_s": 191.26944897628422,
      "api_tail_tok_s": 192.05989866506567,
      "api_ttft_s": 0.638121159048751,
      "api_e2e_s": 0.9666030909866095,
      "draft_accept_rate": 0.9365079365079365,
      "decode_min": 191.1481825351695,
      "decode_max": 191.43350326099483
    },
    {
      "engine": "GGZ14 MXFP4 + DFlash2-7",
      "kind": "pp0",
      "case": "pp8192",
      "target_tokens": 8192,
      "input_tokens": 8077,
      "output_tokens": 64,
      "computed_prompt_tokens": 8077.0,
      "prefill_tok_s": 3345.3259811159983,
      "decode_tok_s": 121.94876919899711,
      "api_tail_tok_s": 122.97040038545589,
      "api_ttft_s": 2.422477585962042,
      "api_e2e_s": 2.9349016379565,
      "draft_accept_rate": 0.5,
      "decode_min": 121.94285513525067,
      "decode_max": 121.96437588108253
    },
    {
      "engine": "GGZ14 MXFP4 + DFlash2-7",
      "kind": "real",
      "case": "real-design",
      "target_tokens": null,
      "input_tokens": 73,
      "output_tokens": 1200,
      "computed_prompt_tokens": 73.0,
      "prefill_tok_s": 859.4747273717237,
      "decode_tok_s": 115.4460927438247,
      "api_tail_tok_s": 115.44592065936577,
      "api_ttft_s": 0.08754590898752213,
      "api_e2e_s": 10.473471021046862,
      "draft_accept_rate": 0.4431984397854705,
      "decode_min": 113.09719122288278,
      "decode_max": 140.28281554922978
    },
    {
      "engine": "GGZ14 MXFP4 + DFlash2-7",
      "kind": "real",
      "case": "real-review",
      "target_tokens": null,
      "input_tokens": 58,
      "output_tokens": 1200,
      "computed_prompt_tokens": 58.0,
      "prefill_tok_s": 1139.6624495876465,
      "decode_tok_s": 68.03603338903416,
      "api_tail_tok_s": 68.03635591112902,
      "api_ttft_s": 0.05375650501810014,
      "api_e2e_s": 17.676777123007923,
      "draft_accept_rate": 0.20276497695852536,
      "decode_min": 66.04682424592336,
      "decode_max": 68.19481406909168
    },
    {
      "engine": "GGZ14 MXFP4 + DFlash2-7",
      "kind": "real",
      "case": "real-qa",
      "target_tokens": null,
      "input_tokens": 61,
      "output_tokens": 1200,
      "computed_prompt_tokens": 61.0,
      "prefill_tok_s": 1194.3088653712255,
      "decode_tok_s": 89.36608597875215,
      "api_tail_tok_s": 89.3678924577195,
      "api_ttft_s": 0.0538298359606415,
      "api_e2e_s": 13.47035250486806,
      "draft_accept_rate": 0.31216931216931215,
      "decode_min": 88.64324746995352,
      "decode_max": 94.30329820507566
    }
  ],
  "telemetry": {
    "llama": [
      {
        "samples": 206,
        "covered_request_s": 205.9000000001397,
        "ppt_average_w": 290.71934113218424,
        "ppt_energy_j": 59859.11233915735,
        "ppt_max_w": 316.0,
        "vram_peak_gib": 15.763111114501953,
        "edge_max_c": 73.0,
        "junction_max_c": 97.0,
        "memory_max_c": 90.0,
        "sclk_median_mhz": 2989.5,
        "mclk_median_mhz": 1258.0,
        "scope": "ladder and real-prompt request windows, including prefill; no warmup/load/pp0"
      },
      {
        "samples": 208,
        "covered_request_s": 207.89999999967404,
        "ppt_average_w": 290.6039904297689,
        "ppt_energy_j": 60416.56961025423,
        "ppt_max_w": 309.0,
        "vram_peak_gib": 15.763736724853516,
        "edge_max_c": 73.0,
        "junction_max_c": 97.0,
        "memory_max_c": 90.0,
        "sclk_median_mhz": 2984.5,
        "mclk_median_mhz": 1258.0,
        "scope": "ladder and real-prompt request windows, including prefill; no warmup/load/pp0"
      },
      {
        "samples": 204,
        "covered_request_s": 205.00000000023283,
        "ppt_average_w": 291.62575344521747,
        "ppt_energy_j": 59783.27945633748,
        "ppt_max_w": 308.0,
        "vram_peak_gib": 15.76260757446289,
        "edge_max_c": 73.0,
        "junction_max_c": 97.0,
        "memory_max_c": 90.0,
        "sclk_median_mhz": 2986.5,
        "mclk_median_mhz": 1258.0,
        "scope": "ladder and real-prompt request windows, including prefill; no warmup/load/pp0"
      }
    ],
    "ggz14": [
      {
        "samples": 92,
        "covered_request_s": 92.11764640524052,
        "ppt_average_w": 290.2690738067196,
        "ppt_energy_j": 26738.903903304057,
        "ppt_max_w": 302.0,
        "vram_peak_gib": 29.417499542236328,
        "edge_max_c": 72.0,
        "junction_max_c": 93.0,
        "memory_max_c": 90.0,
        "sclk_median_mhz": 3216.0,
        "mclk_median_mhz": 1258.0,
        "scope": "ladder and real-prompt request windows, including prefill; no warmup/load/pp0"
      },
      {
        "samples": 87,
        "covered_request_s": 86.71620270586573,
        "ppt_average_w": 298.3085572416642,
        "ppt_energy_j": 25868.1853186625,
        "ppt_max_w": 301.0,
        "vram_peak_gib": 29.417499542236328,
        "edge_max_c": 73.0,
        "junction_max_c": 95.0,
        "memory_max_c": 92.0,
        "sclk_median_mhz": 3199.0,
        "mclk_median_mhz": 1258.0,
        "scope": "ladder and real-prompt request windows, including prefill; no warmup/load/pp0"
      },
      {
        "samples": 88,
        "covered_request_s": 88.62413555383682,
        "ppt_average_w": 298.558607089974,
        "ppt_energy_j": 26459.49846550656,
        "ppt_max_w": 304.0,
        "vram_peak_gib": 29.417499542236328,
        "edge_max_c": 73.0,
        "junction_max_c": 95.0,
        "memory_max_c": 93.0,
        "sclk_median_mhz": 3198.0,
        "mclk_median_mhz": 1258.0,
        "scope": "ladder and real-prompt request windows, including prefill; no warmup/load/pp0"
      }
    ]
  },
  "llama_repetitions": [
    6,
    7,
    8
  ],
  "timing_limitations": [
    "engine timer boundaries differ; configurations include different speculation",
    "llama telemetry windows reconstructed from stdout receipt minus rounded wall_s",
    "vLLM telemetry uses exact client request window",
    "PPT is GPU telemetry, not wall-plug power"
  ]
}