{
  "version": "1",
  "updated": "2026-10-19",
  "data_version": "2026-10-19",
  "unit": "Wh per million output tokens",
  "itl_deadline_ms": 500,
  "source": {
    "name": "ML.ENERGY Leaderboard",
    "publisher": "The ML.ENERGY Initiative (started at SymbioticLab, University of Michigan)",
    "url": "https://ml.energy/leaderboard/",
    "data_url": "https://raw.githubusercontent.com/ml-energy/leaderboard/master/public/data/tasks/lm-arena-chat.json",
    "version": "3.0 (released 2025-12-01); data compiled 2026-02-16",
    "released": "2025-12-01",
    "compiled": "2026-02-16",
    "compiled_evidence": "the leaderboard's public/data/index.json last_updated, and the last commit to the task files, both 2026-02-16",
    "retrieved": "2026-09-26",
    "licence": "Apache-2.0 per the ml-energy/benchmark-v3 dataset card",
    "licence_evidence": [
      "Hugging Face dataset ml-energy/benchmark-v3: tagged license:apache-2.0, ships a LICENSE file, and its dataset card's License section reads Apache-2.0",
      "The ML.ENERGY Leaderboard web app's footer says Apache 2.0",
      "The leaderboard repository (github.com/ml-energy/leaderboard) carries no licence file; its JSON files are reproduced here as published measurements until the ML.ENERGY team confirms that the dataset licence covers them"
    ],
    "workload": "Text Conversation (lm-arena-chat)",
    "method": "Energy per output token is steady-state GPU energy divided by the output tokens generated in the steady-state window (the benchmark's steady_state_energy_per_token), measured with Zeus under vLLM 0.11.1 on H100 80GB HBM3 and B200 servers. Energy per request is that figure multiplied by the average output length, not a separate measurement. Prompt processing is included.",
    "selection": "Each index row is the lowest-energy configuration for its model and GPU whose median inter-token latency does not exceed 500 ms, the leaderboard's own default ceiling; no measured configuration reaches it.",
    "cite": "Chung et al., The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization, NeurIPS 2025 Datasets and Benchmarks, arXiv 2505.06371",
    "cite_url": "https://arxiv.org/abs/2505.06371",
    "used_for": "index, per-model configurations, concurrency curves, measured GPU power",
    "configurations": "328 stable configurations: the leaderboard publishes the runs that passed its stability filter; every run, including the unstable ones, is in the gated Hugging Face dataset ml-energy/benchmark-v3"
  },
  "caveat": "GPU energy only, measured at steady state on a busy server (the most energy-efficient configuration under a latency limit). Host, network, cooling and idle time are not counted. Every figure is a floor; a lightly used server sits well above it.",
  "boundary": "gpu_only",
  "slug": "llama-3-1-405b-instruct-fp8",
  "name": "Llama 3.1 405B Instruct FP8",
  "model_id": "meta-llama/Llama-3.1-405B-Instruct-FP8",
  "params_b": 405.0,
  "active_params_b": 405.0,
  "architecture": "Dense Transformer",
  "gpus": [
    "B200",
    "H100"
  ],
  "configurations": [
    {
      "gpu": "B200",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 8,
      "avg_requests_in_flight": 8.0,
      "wh_per_million_output_tokens": 3607.03,
      "joules_per_output_token": 12.9853248,
      "median_itl_ms": 29.2,
      "p90_itl_ms": 29.7,
      "p99_itl_ms": 41.3,
      "tokens_per_second": 265.8,
      "gpu_power_watts_total": 3451,
      "avg_output_tokens": 376,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 30.0,
      "row_id": "llama-3-1-405b-instruct-fp8/B200/4/8",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 16,
      "avg_requests_in_flight": 15.9,
      "wh_per_million_output_tokens": 1965.46,
      "joules_per_output_token": 7.0756457,
      "median_itl_ms": 31.6,
      "p90_itl_ms": 32.4,
      "p99_itl_ms": 70.3,
      "tokens_per_second": 479.5,
      "gpu_power_watts_total": 3393,
      "avg_output_tokens": 359,
      "under_itl_deadline": true,
      "avg_running_batch": 15.9,
      "p95_itl_ms": 34.0,
      "row_id": "llama-3-1-405b-instruct-fp8/B200/4/16",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 31.9,
      "wh_per_million_output_tokens": 1137.71,
      "joules_per_output_token": 4.0957469,
      "median_itl_ms": 33.7,
      "p90_itl_ms": 36.2,
      "p99_itl_ms": 108.7,
      "tokens_per_second": 865.4,
      "gpu_power_watts_total": 3544,
      "avg_output_tokens": 362,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 42.2,
      "row_id": "llama-3-1-405b-instruct-fp8/B200/4/32",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.9,
      "wh_per_million_output_tokens": 711.48,
      "joules_per_output_token": 2.5613156,
      "median_itl_ms": 38.2,
      "p90_itl_ms": 42.6,
      "p99_itl_ms": 169.7,
      "tokens_per_second": 1434.6,
      "gpu_power_watts_total": 3674,
      "avg_output_tokens": 372,
      "under_itl_deadline": true,
      "avg_running_batch": 63.9,
      "p95_itl_ms": 58.8,
      "row_id": "llama-3-1-405b-instruct-fp8/B200/4/64",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.7,
      "wh_per_million_output_tokens": 492.75,
      "joules_per_output_token": 1.7738933,
      "median_itl_ms": 45.8,
      "p90_itl_ms": 61.6,
      "p99_itl_ms": 250.5,
      "tokens_per_second": 2170.3,
      "gpu_power_watts_total": 3850,
      "avg_output_tokens": 369,
      "under_itl_deadline": true,
      "avg_running_batch": 127.7,
      "p95_itl_ms": 110.8,
      "row_id": "llama-3-1-405b-instruct-fp8/B200/4/128",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.4,
      "wh_per_million_output_tokens": 376.25,
      "joules_per_output_token": 1.3544945,
      "median_itl_ms": 63.4,
      "p90_itl_ms": 115.1,
      "p99_itl_ms": 431.8,
      "tokens_per_second": 2878.1,
      "gpu_power_watts_total": 3898,
      "avg_output_tokens": 375,
      "under_itl_deadline": true,
      "avg_running_batch": 255.4,
      "p95_itl_ms": 151.1,
      "row_id": "llama-3-1-405b-instruct-fp8/B200/4/256",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 512,
      "avg_requests_in_flight": 510.5,
      "wh_per_million_output_tokens": 318.53,
      "joules_per_output_token": 1.1466946,
      "median_itl_ms": 98.2,
      "p90_itl_ms": 186.1,
      "p99_itl_ms": 959.1,
      "tokens_per_second": 3386.3,
      "gpu_power_watts_total": 3883,
      "avg_output_tokens": 380,
      "under_itl_deadline": true,
      "avg_running_batch": 510.5,
      "p95_itl_ms": 248.8,
      "row_id": "llama-3-1-405b-instruct-fp8/B200/4/512",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 8,
      "avg_requests_in_flight": 8.0,
      "wh_per_million_output_tokens": 3881.91,
      "joules_per_output_token": 13.9748687,
      "median_itl_ms": 33.2,
      "p90_itl_ms": 33.5,
      "p99_itl_ms": 42.5,
      "tokens_per_second": 233.3,
      "gpu_power_watts_total": 3260,
      "avg_output_tokens": 366,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 33.6,
      "row_id": "llama-3-1-405b-instruct-fp8/H100/8/8",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 16,
      "avg_requests_in_flight": 15.9,
      "wh_per_million_output_tokens": 2108.12,
      "joules_per_output_token": 7.5892304,
      "median_itl_ms": 34.0,
      "p90_itl_ms": 34.3,
      "p99_itl_ms": 85.4,
      "tokens_per_second": 445.9,
      "gpu_power_watts_total": 3384,
      "avg_output_tokens": 380,
      "under_itl_deadline": true,
      "avg_running_batch": 15.9,
      "p95_itl_ms": 34.7,
      "row_id": "llama-3-1-405b-instruct-fp8/H100/8/16",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 31.9,
      "wh_per_million_output_tokens": 1323.69,
      "joules_per_output_token": 4.7652848,
      "median_itl_ms": 36.3,
      "p90_itl_ms": 36.9,
      "p99_itl_ms": 137.0,
      "tokens_per_second": 785.7,
      "gpu_power_watts_total": 3744,
      "avg_output_tokens": 372,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 43.7,
      "row_id": "llama-3-1-405b-instruct-fp8/H100/8/32",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.9,
      "wh_per_million_output_tokens": 856.24,
      "joules_per_output_token": 3.082474,
      "median_itl_ms": 39.3,
      "p90_itl_ms": 43.9,
      "p99_itl_ms": 190.2,
      "tokens_per_second": 1350.2,
      "gpu_power_watts_total": 4162,
      "avg_output_tokens": 360,
      "under_itl_deadline": true,
      "avg_running_batch": 63.9,
      "p95_itl_ms": 92.2,
      "row_id": "llama-3-1-405b-instruct-fp8/H100/8/64",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 96,
      "avg_requests_in_flight": 95.7,
      "wh_per_million_output_tokens": 739.14,
      "joules_per_output_token": 2.6609122,
      "median_itl_ms": 44.1,
      "p90_itl_ms": 57.9,
      "p99_itl_ms": 280.6,
      "tokens_per_second": 1681.6,
      "gpu_power_watts_total": 4475,
      "avg_output_tokens": 368,
      "under_itl_deadline": true,
      "avg_running_batch": 95.7,
      "p95_itl_ms": 104.2,
      "row_id": "llama-3-1-405b-instruct-fp8/H100/8/96",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.7,
      "wh_per_million_output_tokens": 674.86,
      "joules_per_output_token": 2.4295126,
      "median_itl_ms": 48.0,
      "p90_itl_ms": 99.4,
      "p99_itl_ms": 353.7,
      "tokens_per_second": 1895.9,
      "gpu_power_watts_total": 4606,
      "avg_output_tokens": 361,
      "under_itl_deadline": true,
      "avg_running_batch": 127.7,
      "p95_itl_ms": 142.8,
      "row_id": "llama-3-1-405b-instruct-fp8/H100/8/128",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 192,
      "avg_requests_in_flight": 191.5,
      "wh_per_million_output_tokens": 620.53,
      "joules_per_output_token": 2.2338905,
      "median_itl_ms": 62.2,
      "p90_itl_ms": 113.5,
      "p99_itl_ms": 424.7,
      "tokens_per_second": 2154.5,
      "gpu_power_watts_total": 4813,
      "avg_output_tokens": 359,
      "under_itl_deadline": true,
      "avg_running_batch": 191.5,
      "p95_itl_ms": 182.4,
      "row_id": "llama-3-1-405b-instruct-fp8/H100/8/192",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.2,
      "wh_per_million_output_tokens": 572.94,
      "joules_per_output_token": 2.0625855,
      "median_itl_ms": 66.9,
      "p90_itl_ms": 148.2,
      "p99_itl_ms": 569.3,
      "tokens_per_second": 2363.4,
      "gpu_power_watts_total": 4875,
      "avg_output_tokens": 367,
      "under_itl_deadline": true,
      "avg_running_batch": 255.2,
      "p95_itl_ms": 194.3,
      "row_id": "llama-3-1-405b-instruct-fp8/H100/8/256",
      "repeat": 1
    }
  ],
  "best": {
    "B200": {
      "gpu": "B200",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 512,
      "avg_requests_in_flight": 510.5,
      "wh_per_million_output_tokens": 318.53,
      "joules_per_output_token": 1.1466946,
      "median_itl_ms": 98.2,
      "p90_itl_ms": 186.1,
      "p99_itl_ms": 959.1,
      "tokens_per_second": 3386.3,
      "gpu_power_watts_total": 3883,
      "avg_output_tokens": 380,
      "under_itl_deadline": true,
      "avg_running_batch": 510.5,
      "p95_itl_ms": 248.8,
      "row_id": "llama-3-1-405b-instruct-fp8/B200/4/512",
      "repeat": 1
    },
    "H100": {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "FP8",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.2,
      "wh_per_million_output_tokens": 572.94,
      "joules_per_output_token": 2.0625855,
      "median_itl_ms": 66.9,
      "p90_itl_ms": 148.2,
      "p99_itl_ms": 569.3,
      "tokens_per_second": 2363.4,
      "gpu_power_watts_total": 4875,
      "avg_output_tokens": 367,
      "under_itl_deadline": true,
      "avg_running_batch": 255.2,
      "p95_itl_ms": 194.3,
      "row_id": "llama-3-1-405b-instruct-fp8/H100/8/256",
      "repeat": 1
    }
  }
}