{
  "version": "1",
  "updated": "2026-09-27",
  "data_version": "2026-09-27",
  "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",
  "name": "Llama 3.1 405B Instruct",
  "model_id": "meta-llama/Llama-3.1-405B-Instruct",
  "params_b": 405.0,
  "active_params_b": 405.0,
  "architecture": "Dense Transformer",
  "gpus": [
    "B200"
  ],
  "configurations": [
    {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 8,
      "avg_requests_in_flight": 8.0,
      "wh_per_million_output_tokens": 6073.44,
      "joules_per_output_token": 21.8643741,
      "median_itl_ms": 29.1,
      "p90_itl_ms": 29.4,
      "p99_itl_ms": 37.7,
      "tokens_per_second": 268.8,
      "gpu_power_watts_total": 5878,
      "avg_output_tokens": 363,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 29.7,
      "row_id": "llama-3-1-405b-instruct/B200/8/8",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 16,
      "avg_requests_in_flight": 15.9,
      "wh_per_million_output_tokens": 3228.0,
      "joules_per_output_token": 11.6208115,
      "median_itl_ms": 29.4,
      "p90_itl_ms": 29.9,
      "p99_itl_ms": 51.7,
      "tokens_per_second": 523.0,
      "gpu_power_watts_total": 6077,
      "avg_output_tokens": 368,
      "under_itl_deadline": true,
      "avg_running_batch": 15.9,
      "p95_itl_ms": 30.6,
      "row_id": "llama-3-1-405b-instruct/B200/8/16",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 31.9,
      "wh_per_million_output_tokens": 1777.63,
      "joules_per_output_token": 6.3994752,
      "median_itl_ms": 28.0,
      "p90_itl_ms": 29.5,
      "p99_itl_ms": 81.1,
      "tokens_per_second": 1052.4,
      "gpu_power_watts_total": 6735,
      "avg_output_tokens": 364,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 35.4,
      "row_id": "llama-3-1-405b-instruct/B200/8/32",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.8,
      "wh_per_million_output_tokens": 1133.93,
      "joules_per_output_token": 4.082157,
      "median_itl_ms": 30.8,
      "p90_itl_ms": 34.5,
      "p99_itl_ms": 128.2,
      "tokens_per_second": 1800.4,
      "gpu_power_watts_total": 7350,
      "avg_output_tokens": 365,
      "under_itl_deadline": true,
      "avg_running_batch": 63.8,
      "p95_itl_ms": 48.4,
      "row_id": "llama-3-1-405b-instruct/B200/8/64",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.6,
      "wh_per_million_output_tokens": 750.18,
      "joules_per_output_token": 2.700665,
      "median_itl_ms": 36.5,
      "p90_itl_ms": 46.8,
      "p99_itl_ms": 192.6,
      "tokens_per_second": 2804.9,
      "gpu_power_watts_total": 7575,
      "avg_output_tokens": 372,
      "under_itl_deadline": true,
      "avg_running_batch": 127.6,
      "p95_itl_ms": 81.1,
      "row_id": "llama-3-1-405b-instruct/B200/8/128",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.1,
      "wh_per_million_output_tokens": 591.87,
      "joules_per_output_token": 2.1307191,
      "median_itl_ms": 49.3,
      "p90_itl_ms": 96.0,
      "p99_itl_ms": 319.0,
      "tokens_per_second": 3587.0,
      "gpu_power_watts_total": 7643,
      "avg_output_tokens": 366,
      "under_itl_deadline": true,
      "avg_running_batch": 255.1,
      "p95_itl_ms": 135.1,
      "row_id": "llama-3-1-405b-instruct/B200/8/256",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 512,
      "avg_requests_in_flight": 510.5,
      "wh_per_million_output_tokens": 467.22,
      "joules_per_output_token": 1.681983,
      "median_itl_ms": 77.2,
      "p90_itl_ms": 136.4,
      "p99_itl_ms": 849.9,
      "tokens_per_second": 4550.9,
      "gpu_power_watts_total": 7655,
      "avg_output_tokens": 358,
      "under_itl_deadline": true,
      "avg_running_batch": 510.5,
      "p95_itl_ms": 185.6,
      "row_id": "llama-3-1-405b-instruct/B200/8/512",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 768,
      "avg_requests_in_flight": 765.7,
      "wh_per_million_output_tokens": 379.66,
      "joules_per_output_token": 1.3667623,
      "median_itl_ms": 122.2,
      "p90_itl_ms": 151.1,
      "p99_itl_ms": 851.2,
      "tokens_per_second": 5496.0,
      "gpu_power_watts_total": 7512,
      "avg_output_tokens": 372,
      "under_itl_deadline": true,
      "avg_running_batch": 765.7,
      "p95_itl_ms": 196.6,
      "row_id": "llama-3-1-405b-instruct/B200/8/768",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 1024,
      "avg_requests_in_flight": 1021.3,
      "wh_per_million_output_tokens": 331.46,
      "joules_per_output_token": 1.1932728,
      "median_itl_ms": 146.0,
      "p90_itl_ms": 162.2,
      "p99_itl_ms": 856.4,
      "tokens_per_second": 6461.4,
      "gpu_power_watts_total": 7710,
      "avg_output_tokens": 372,
      "under_itl_deadline": true,
      "avg_running_batch": 1021.3,
      "p95_itl_ms": 172.5,
      "row_id": "llama-3-1-405b-instruct/B200/8/1024",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 2048,
      "avg_requests_in_flight": 2042.8,
      "wh_per_million_output_tokens": 297.42,
      "joules_per_output_token": 1.0706985,
      "median_itl_ms": 239.6,
      "p90_itl_ms": 291.8,
      "p99_itl_ms": 870.0,
      "tokens_per_second": 7173.1,
      "gpu_power_watts_total": 7680,
      "avg_output_tokens": 362,
      "under_itl_deadline": true,
      "avg_running_batch": 2042.8,
      "p95_itl_ms": 303.1,
      "row_id": "llama-3-1-405b-instruct/B200/8/2048",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 3072,
      "avg_requests_in_flight": 3062.9,
      "wh_per_million_output_tokens": 289.8,
      "joules_per_output_token": 1.0432878,
      "median_itl_ms": 325.5,
      "p90_itl_ms": 428.6,
      "p99_itl_ms": 914.0,
      "tokens_per_second": 7173.7,
      "gpu_power_watts_total": 7484,
      "avg_output_tokens": 371,
      "under_itl_deadline": true,
      "avg_running_batch": 3062.9,
      "p95_itl_ms": 465.4,
      "row_id": "llama-3-1-405b-instruct/B200/8/3072",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 4096,
      "avg_requests_in_flight": 3588.7,
      "wh_per_million_output_tokens": 281.96,
      "joules_per_output_token": 1.0150418,
      "median_itl_ms": 272.3,
      "p90_itl_ms": 544.9,
      "p99_itl_ms": 977.2,
      "tokens_per_second": 7418.6,
      "gpu_power_watts_total": 7530,
      "avg_output_tokens": 370,
      "under_itl_deadline": true,
      "avg_running_batch": 3588.7,
      "p95_itl_ms": 616.3,
      "row_id": "llama-3-1-405b-instruct/B200/8/4096",
      "repeat": 1
    }
  ],
  "best": {
    "B200": {
      "gpu": "B200",
      "num_gpus": 8,
      "tensor_parallel": 8,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 4096,
      "avg_requests_in_flight": 3588.7,
      "wh_per_million_output_tokens": 281.96,
      "joules_per_output_token": 1.0150418,
      "median_itl_ms": 272.3,
      "p90_itl_ms": 544.9,
      "p99_itl_ms": 977.2,
      "tokens_per_second": 7418.6,
      "gpu_power_watts_total": 7530,
      "avg_output_tokens": 370,
      "under_itl_deadline": true,
      "avg_running_batch": 3588.7,
      "p95_itl_ms": 616.3,
      "row_id": "llama-3-1-405b-instruct/B200/8/4096",
      "repeat": 1
    }
  }
}