{
  "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-3-3-70b-instruct",
  "name": "Llama 3.3 70B Instruct",
  "model_id": "meta-llama/Llama-3.3-70B-Instruct",
  "params_b": 70.0,
  "active_params_b": 70.0,
  "architecture": "Dense Transformer",
  "gpus": [
    "B200",
    "H100"
  ],
  "configurations": [
    {
      "gpu": "B200",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 8,
      "avg_requests_in_flight": 8.0,
      "wh_per_million_output_tokens": 1004.67,
      "joules_per_output_token": 3.6168273,
      "median_itl_ms": 16.6,
      "p90_itl_ms": 17.1,
      "p99_itl_ms": 21.9,
      "tokens_per_second": 470.4,
      "gpu_power_watts_total": 1701,
      "avg_output_tokens": 411,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 17.4,
      "row_id": "llama-3-3-70b-instruct/B200/2/8",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 16,
      "avg_requests_in_flight": 16.0,
      "wh_per_million_output_tokens": 533.75,
      "joules_per_output_token": 1.9214979,
      "median_itl_ms": 16.9,
      "p90_itl_ms": 17.7,
      "p99_itl_ms": 28.9,
      "tokens_per_second": 908.4,
      "gpu_power_watts_total": 1746,
      "avg_output_tokens": 406,
      "under_itl_deadline": true,
      "avg_running_batch": 16.0,
      "p95_itl_ms": 18.3,
      "row_id": "llama-3-3-70b-instruct/B200/2/16",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 31.9,
      "wh_per_million_output_tokens": 305.1,
      "joules_per_output_token": 1.0983761,
      "median_itl_ms": 18.6,
      "p90_itl_ms": 20.2,
      "p99_itl_ms": 46.6,
      "tokens_per_second": 1600.2,
      "gpu_power_watts_total": 1758,
      "avg_output_tokens": 407,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 23.3,
      "row_id": "llama-3-3-70b-instruct/B200/2/32",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.8,
      "wh_per_million_output_tokens": 189.89,
      "joules_per_output_token": 0.6835897,
      "median_itl_ms": 21.8,
      "p90_itl_ms": 24.8,
      "p99_itl_ms": 69.5,
      "tokens_per_second": 2655.3,
      "gpu_power_watts_total": 1815,
      "avg_output_tokens": 415,
      "under_itl_deadline": true,
      "avg_running_batch": 63.8,
      "p95_itl_ms": 31.7,
      "row_id": "llama-3-3-70b-instruct/B200/2/64",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.7,
      "wh_per_million_output_tokens": 133.64,
      "joules_per_output_token": 0.4810869,
      "median_itl_ms": 27.6,
      "p90_itl_ms": 33.6,
      "p99_itl_ms": 107.7,
      "tokens_per_second": 3973.4,
      "gpu_power_watts_total": 1912,
      "avg_output_tokens": 410,
      "under_itl_deadline": true,
      "avg_running_batch": 127.7,
      "p95_itl_ms": 50.7,
      "row_id": "llama-3-3-70b-instruct/B200/2/128",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.5,
      "wh_per_million_output_tokens": 97.82,
      "joules_per_output_token": 0.3521439,
      "median_itl_ms": 36.7,
      "p90_itl_ms": 56.2,
      "p99_itl_ms": 170.3,
      "tokens_per_second": 5515.2,
      "gpu_power_watts_total": 1942,
      "avg_output_tokens": 410,
      "under_itl_deadline": true,
      "avg_running_batch": 255.5,
      "p95_itl_ms": 78.2,
      "row_id": "llama-3-3-70b-instruct/B200/2/256",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 512,
      "avg_requests_in_flight": 510.6,
      "wh_per_million_output_tokens": 75.74,
      "joules_per_output_token": 0.2726697,
      "median_itl_ms": 63.5,
      "p90_itl_ms": 83.9,
      "p99_itl_ms": 183.9,
      "tokens_per_second": 7078.6,
      "gpu_power_watts_total": 1930,
      "avg_output_tokens": 414,
      "under_itl_deadline": true,
      "avg_running_batch": 510.6,
      "p95_itl_ms": 104.1,
      "row_id": "llama-3-3-70b-instruct/B200/2/512",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 1024,
      "avg_requests_in_flight": 1021.6,
      "wh_per_million_output_tokens": 60.72,
      "joules_per_output_token": 0.218586,
      "median_itl_ms": 105.6,
      "p90_itl_ms": 123.8,
      "p99_itl_ms": 514.4,
      "tokens_per_second": 8853.9,
      "gpu_power_watts_total": 1935,
      "avg_output_tokens": 409,
      "under_itl_deadline": true,
      "avg_running_batch": 1021.6,
      "p95_itl_ms": 138.3,
      "row_id": "llama-3-3-70b-instruct/B200/2/1024",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 8,
      "avg_requests_in_flight": 8.0,
      "wh_per_million_output_tokens": 1068.59,
      "joules_per_output_token": 3.8469267,
      "median_itl_ms": 16.9,
      "p90_itl_ms": 17.0,
      "p99_itl_ms": 20.6,
      "tokens_per_second": 464.6,
      "gpu_power_watts_total": 1787,
      "avg_output_tokens": 409,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 17.1,
      "row_id": "llama-3-3-70b-instruct/H100/4/8",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 16,
      "avg_requests_in_flight": 16.0,
      "wh_per_million_output_tokens": 583.79,
      "joules_per_output_token": 2.1016443,
      "median_itl_ms": 17.6,
      "p90_itl_ms": 17.9,
      "p99_itl_ms": 28.5,
      "tokens_per_second": 874.9,
      "gpu_power_watts_total": 1839,
      "avg_output_tokens": 410,
      "under_itl_deadline": true,
      "avg_running_batch": 16.0,
      "p95_itl_ms": 18.2,
      "row_id": "llama-3-3-70b-instruct/H100/4/16",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 31.9,
      "wh_per_million_output_tokens": 346.96,
      "joules_per_output_token": 1.2490713,
      "median_itl_ms": 19.9,
      "p90_itl_ms": 20.4,
      "p99_itl_ms": 51.3,
      "tokens_per_second": 1517.6,
      "gpu_power_watts_total": 1896,
      "avg_output_tokens": 413,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 22.0,
      "row_id": "llama-3-3-70b-instruct/H100/4/32",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.9,
      "wh_per_million_output_tokens": 228.42,
      "joules_per_output_token": 0.8223267,
      "median_itl_ms": 22.3,
      "p90_itl_ms": 23.9,
      "p99_itl_ms": 85.0,
      "tokens_per_second": 2563.8,
      "gpu_power_watts_total": 2108,
      "avg_output_tokens": 409,
      "under_itl_deadline": true,
      "avg_running_batch": 63.9,
      "p95_itl_ms": 29.1,
      "row_id": "llama-3-3-70b-instruct/H100/4/64",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 96,
      "avg_requests_in_flight": 95.9,
      "wh_per_million_output_tokens": 186.32,
      "joules_per_output_token": 0.6707632,
      "median_itl_ms": 25.4,
      "p90_itl_ms": 27.7,
      "p99_itl_ms": 100.2,
      "tokens_per_second": 3315.3,
      "gpu_power_watts_total": 2224,
      "avg_output_tokens": 420,
      "under_itl_deadline": true,
      "avg_running_batch": 95.9,
      "p95_itl_ms": 39.7,
      "row_id": "llama-3-3-70b-instruct/H100/4/96",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.7,
      "wh_per_million_output_tokens": 167.67,
      "joules_per_output_token": 0.6036207,
      "median_itl_ms": 27.8,
      "p90_itl_ms": 32.1,
      "p99_itl_ms": 123.1,
      "tokens_per_second": 3895.4,
      "gpu_power_watts_total": 2351,
      "avg_output_tokens": 411,
      "under_itl_deadline": true,
      "avg_running_batch": 127.7,
      "p95_itl_ms": 54.1,
      "row_id": "llama-3-3-70b-instruct/H100/4/128",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 192,
      "avg_requests_in_flight": 191.5,
      "wh_per_million_output_tokens": 150.0,
      "joules_per_output_token": 0.5400162,
      "median_itl_ms": 33.6,
      "p90_itl_ms": 48.6,
      "p99_itl_ms": 171.2,
      "tokens_per_second": 4557.9,
      "gpu_power_watts_total": 2461,
      "avg_output_tokens": 406,
      "under_itl_deadline": true,
      "avg_running_batch": 191.5,
      "p95_itl_ms": 75.2,
      "row_id": "llama-3-3-70b-instruct/H100/4/192",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.5,
      "wh_per_million_output_tokens": 136.41,
      "joules_per_output_token": 0.491087,
      "median_itl_ms": 39.1,
      "p90_itl_ms": 59.9,
      "p99_itl_ms": 181.4,
      "tokens_per_second": 5130.2,
      "gpu_power_watts_total": 2519,
      "avg_output_tokens": 413,
      "under_itl_deadline": true,
      "avg_running_batch": 255.5,
      "p95_itl_ms": 85.3,
      "row_id": "llama-3-3-70b-instruct/H100/4/256",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 384,
      "avg_requests_in_flight": 382.9,
      "wh_per_million_output_tokens": 123.26,
      "joules_per_output_token": 0.4437462,
      "median_itl_ms": 53.8,
      "p90_itl_ms": 81.2,
      "p99_itl_ms": 268.9,
      "tokens_per_second": 5657.8,
      "gpu_power_watts_total": 2511,
      "avg_output_tokens": 408,
      "under_itl_deadline": true,
      "avg_running_batch": 382.9,
      "p95_itl_ms": 104.6,
      "row_id": "llama-3-3-70b-instruct/H100/4/384",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 512,
      "avg_requests_in_flight": 510.8,
      "wh_per_million_output_tokens": 118.09,
      "joules_per_output_token": 0.4251173,
      "median_itl_ms": 73.6,
      "p90_itl_ms": 104.2,
      "p99_itl_ms": 277.3,
      "tokens_per_second": 5858.5,
      "gpu_power_watts_total": 2491,
      "avg_output_tokens": 414,
      "under_itl_deadline": true,
      "avg_running_batch": 510.8,
      "p95_itl_ms": 134.7,
      "row_id": "llama-3-3-70b-instruct/H100/4/512",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 768,
      "avg_requests_in_flight": 724.2,
      "wh_per_million_output_tokens": 117.53,
      "joules_per_output_token": 0.4230942,
      "median_itl_ms": 97.5,
      "p90_itl_ms": 149.5,
      "p99_itl_ms": 570.3,
      "tokens_per_second": 6017.7,
      "gpu_power_watts_total": 2546,
      "avg_output_tokens": 413,
      "under_itl_deadline": true,
      "avg_running_batch": 724.2,
      "p95_itl_ms": 193.7,
      "row_id": "llama-3-3-70b-instruct/H100/4/768",
      "repeat": 1
    }
  ],
  "best": {
    "B200": {
      "gpu": "B200",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 1024,
      "avg_requests_in_flight": 1021.6,
      "wh_per_million_output_tokens": 60.72,
      "joules_per_output_token": 0.218586,
      "median_itl_ms": 105.6,
      "p90_itl_ms": 123.8,
      "p99_itl_ms": 514.4,
      "tokens_per_second": 8853.9,
      "gpu_power_watts_total": 1935,
      "avg_output_tokens": 409,
      "under_itl_deadline": true,
      "avg_running_batch": 1021.6,
      "p95_itl_ms": 138.3,
      "row_id": "llama-3-3-70b-instruct/B200/2/1024",
      "repeat": 1
    },
    "H100": {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 4,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 768,
      "avg_requests_in_flight": 724.2,
      "wh_per_million_output_tokens": 117.53,
      "joules_per_output_token": 0.4230942,
      "median_itl_ms": 97.5,
      "p90_itl_ms": 149.5,
      "p99_itl_ms": 570.3,
      "tokens_per_second": 6017.7,
      "gpu_power_watts_total": 2546,
      "avg_output_tokens": 413,
      "under_itl_deadline": true,
      "avg_running_batch": 724.2,
      "p95_itl_ms": 193.7,
      "row_id": "llama-3-3-70b-instruct/H100/4/768",
      "repeat": 1
    }
  }
}