{
  "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-4-maverick-17b-128e-instruct-fp8",
  "name": "Llama 4 Maverick 17B 128E Instruct FP8",
  "model_id": "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
  "params_b": 400.0,
  "active_params_b": 17.0,
  "architecture": "MoE",
  "gpus": [
    "H100"
  ],
  "configurations": [
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "FP8",
      "max_num_seqs": 8,
      "avg_requests_in_flight": 8.0,
      "wh_per_million_output_tokens": 779.04,
      "joules_per_output_token": 2.8045326,
      "median_itl_ms": 11.5,
      "p90_itl_ms": 11.7,
      "p99_itl_ms": 46.8,
      "tokens_per_second": 658.0,
      "gpu_power_watts_total": 1845,
      "avg_output_tokens": 415,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 11.8,
      "row_id": "llama-4-maverick-17b-128e-instruct-fp8/H100/8/8",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "FP8",
      "max_num_seqs": 16,
      "avg_requests_in_flight": 16.0,
      "wh_per_million_output_tokens": 504.39,
      "joules_per_output_token": 1.8157943,
      "median_itl_ms": 13.3,
      "p90_itl_ms": 13.6,
      "p99_itl_ms": 48.6,
      "tokens_per_second": 1109.5,
      "gpu_power_watts_total": 2015,
      "avg_output_tokens": 418,
      "under_itl_deadline": true,
      "avg_running_batch": 16.0,
      "p95_itl_ms": 13.9,
      "row_id": "llama-4-maverick-17b-128e-instruct-fp8/H100/8/16",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "FP8",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 31.9,
      "wh_per_million_output_tokens": 351.54,
      "joules_per_output_token": 1.2655324,
      "median_itl_ms": 16.1,
      "p90_itl_ms": 16.7,
      "p99_itl_ms": 50.5,
      "tokens_per_second": 1805.0,
      "gpu_power_watts_total": 2284,
      "avg_output_tokens": 412,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 18.2,
      "row_id": "llama-4-maverick-17b-128e-instruct-fp8/H100/8/32",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "FP8",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.9,
      "wh_per_million_output_tokens": 262.2,
      "joules_per_output_token": 0.9439309,
      "median_itl_ms": 20.7,
      "p90_itl_ms": 22.4,
      "p99_itl_ms": 53.1,
      "tokens_per_second": 2813.4,
      "gpu_power_watts_total": 2656,
      "avg_output_tokens": 406,
      "under_itl_deadline": true,
      "avg_running_batch": 63.9,
      "p95_itl_ms": 47.6,
      "row_id": "llama-4-maverick-17b-128e-instruct-fp8/H100/8/64",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "FP8",
      "max_num_seqs": 96,
      "avg_requests_in_flight": 95.8,
      "wh_per_million_output_tokens": 235.77,
      "joules_per_output_token": 0.8487846,
      "median_itl_ms": 24.5,
      "p90_itl_ms": 26.6,
      "p99_itl_ms": 56.7,
      "tokens_per_second": 3546.0,
      "gpu_power_watts_total": 3010,
      "avg_output_tokens": 408,
      "under_itl_deadline": true,
      "avg_running_batch": 95.8,
      "p95_itl_ms": 50.3,
      "row_id": "llama-4-maverick-17b-128e-instruct-fp8/H100/8/96",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "FP8",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.7,
      "wh_per_million_output_tokens": 210.37,
      "joules_per_output_token": 0.7573397,
      "median_itl_ms": 27.6,
      "p90_itl_ms": 31.3,
      "p99_itl_ms": 59.4,
      "tokens_per_second": 4146.8,
      "gpu_power_watts_total": 3141,
      "avg_output_tokens": 406,
      "under_itl_deadline": true,
      "avg_running_batch": 127.7,
      "p95_itl_ms": 52.7,
      "row_id": "llama-4-maverick-17b-128e-instruct-fp8/H100/8/128",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "FP8",
      "max_num_seqs": 192,
      "avg_requests_in_flight": 191.6,
      "wh_per_million_output_tokens": 178.22,
      "joules_per_output_token": 0.6416027,
      "median_itl_ms": 33.2,
      "p90_itl_ms": 38.1,
      "p99_itl_ms": 66.6,
      "tokens_per_second": 5210.4,
      "gpu_power_watts_total": 3343,
      "avg_output_tokens": 415,
      "under_itl_deadline": true,
      "avg_running_batch": 191.6,
      "p95_itl_ms": 56.6,
      "row_id": "llama-4-maverick-17b-128e-instruct-fp8/H100/8/192",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "FP8",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.4,
      "wh_per_million_output_tokens": 155.69,
      "joules_per_output_token": 0.5604892,
      "median_itl_ms": 37.0,
      "p90_itl_ms": 41.9,
      "p99_itl_ms": 84.0,
      "tokens_per_second": 6257.6,
      "gpu_power_watts_total": 3507,
      "avg_output_tokens": 409,
      "under_itl_deadline": true,
      "avg_running_batch": 255.4,
      "p95_itl_ms": 60.6,
      "row_id": "llama-4-maverick-17b-128e-instruct-fp8/H100/8/256",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "FP8",
      "max_num_seqs": 384,
      "avg_requests_in_flight": 383.3,
      "wh_per_million_output_tokens": 134.5,
      "joules_per_output_token": 0.4841928,
      "median_itl_ms": 47.9,
      "p90_itl_ms": 66.2,
      "p99_itl_ms": 112.5,
      "tokens_per_second": 7149.1,
      "gpu_power_watts_total": 3462,
      "avg_output_tokens": 408,
      "under_itl_deadline": true,
      "avg_running_batch": 383.3,
      "p95_itl_ms": 69.3,
      "row_id": "llama-4-maverick-17b-128e-instruct-fp8/H100/8/384",
      "repeat": 1
    }
  ],
  "best": {
    "H100": {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "FP8",
      "max_num_seqs": 384,
      "avg_requests_in_flight": 383.3,
      "wh_per_million_output_tokens": 134.5,
      "joules_per_output_token": 0.4841928,
      "median_itl_ms": 47.9,
      "p90_itl_ms": 66.2,
      "p99_itl_ms": 112.5,
      "tokens_per_second": 7149.1,
      "gpu_power_watts_total": 3462,
      "avg_output_tokens": 408,
      "under_itl_deadline": true,
      "avg_running_batch": 383.3,
      "p95_itl_ms": 69.3,
      "row_id": "llama-4-maverick-17b-128e-instruct-fp8/H100/8/384",
      "repeat": 1
    }
  }
}