{
  "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-scout-17b-16e-instruct",
  "name": "Llama 4 Scout 17B 16E Instruct",
  "model_id": "meta-llama/Llama-4-Scout-17B-16E-Instruct",
  "params_b": 109.0,
  "active_params_b": 17.0,
  "architecture": "MoE",
  "gpus": [
    "H100"
  ],
  "configurations": [
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 4,
      "precision": "BF16",
      "max_num_seqs": 8,
      "avg_requests_in_flight": 8.0,
      "wh_per_million_output_tokens": 1104.8,
      "joules_per_output_token": 3.977293,
      "median_itl_ms": 24.2,
      "p90_itl_ms": 24.9,
      "p99_itl_ms": 55.3,
      "tokens_per_second": 322.0,
      "gpu_power_watts_total": 1281,
      "avg_output_tokens": 396,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 25.2,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/4/8",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 4,
      "precision": "BF16",
      "max_num_seqs": 16,
      "avg_requests_in_flight": 16.0,
      "wh_per_million_output_tokens": 742.62,
      "joules_per_output_token": 2.6734328,
      "median_itl_ms": 29.0,
      "p90_itl_ms": 29.7,
      "p99_itl_ms": 56.5,
      "tokens_per_second": 532.3,
      "gpu_power_watts_total": 1423,
      "avg_output_tokens": 403,
      "under_itl_deadline": true,
      "avg_running_batch": 16.0,
      "p95_itl_ms": 30.1,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/4/16",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 4,
      "precision": "BF16",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 31.9,
      "wh_per_million_output_tokens": 480.3,
      "joules_per_output_token": 1.7290925,
      "median_itl_ms": 33.0,
      "p90_itl_ms": 33.7,
      "p99_itl_ms": 59.3,
      "tokens_per_second": 928.8,
      "gpu_power_watts_total": 1606,
      "avg_output_tokens": 401,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 34.8,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/4/32",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 4,
      "precision": "BF16",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.9,
      "wh_per_million_output_tokens": 294.26,
      "joules_per_output_token": 1.0593195,
      "median_itl_ms": 36.5,
      "p90_itl_ms": 37.5,
      "p99_itl_ms": 69.6,
      "tokens_per_second": 1651.6,
      "gpu_power_watts_total": 1750,
      "avg_output_tokens": 404,
      "under_itl_deadline": true,
      "avg_running_batch": 63.9,
      "p95_itl_ms": 54.0,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/4/64",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 4,
      "precision": "BF16",
      "max_num_seqs": 96,
      "avg_requests_in_flight": 95.7,
      "wh_per_million_output_tokens": 220.49,
      "joules_per_output_token": 0.7937582,
      "median_itl_ms": 39.0,
      "p90_itl_ms": 41.1,
      "p99_itl_ms": 79.9,
      "tokens_per_second": 2292.2,
      "gpu_power_watts_total": 1819,
      "avg_output_tokens": 410,
      "under_itl_deadline": true,
      "avg_running_batch": 95.7,
      "p95_itl_ms": 58.4,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/4/96",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 4,
      "precision": "BF16",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.6,
      "wh_per_million_output_tokens": 181.66,
      "joules_per_output_token": 0.653973,
      "median_itl_ms": 42.1,
      "p90_itl_ms": 45.7,
      "p99_itl_ms": 94.8,
      "tokens_per_second": 2805.6,
      "gpu_power_watts_total": 1835,
      "avg_output_tokens": 403,
      "under_itl_deadline": true,
      "avg_running_batch": 127.6,
      "p95_itl_ms": 60.9,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/4/128",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 4,
      "precision": "BF16",
      "max_num_seqs": 192,
      "avg_requests_in_flight": 191.6,
      "wh_per_million_output_tokens": 143.44,
      "joules_per_output_token": 0.5163722,
      "median_itl_ms": 47.6,
      "p90_itl_ms": 55.7,
      "p99_itl_ms": 124.2,
      "tokens_per_second": 3614.8,
      "gpu_power_watts_total": 1867,
      "avg_output_tokens": 414,
      "under_itl_deadline": true,
      "avg_running_batch": 191.6,
      "p95_itl_ms": 72.5,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/4/192",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 4,
      "precision": "BF16",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.3,
      "wh_per_million_output_tokens": 120.69,
      "joules_per_output_token": 0.4344742,
      "median_itl_ms": 52.1,
      "p90_itl_ms": 69.9,
      "p99_itl_ms": 158.1,
      "tokens_per_second": 4325.1,
      "gpu_power_watts_total": 1879,
      "avg_output_tokens": 398,
      "under_itl_deadline": true,
      "avg_running_batch": 255.3,
      "p95_itl_ms": 82.7,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/4/256",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 4,
      "precision": "BF16",
      "max_num_seqs": 384,
      "avg_requests_in_flight": 383.1,
      "wh_per_million_output_tokens": 99.7,
      "joules_per_output_token": 0.3589103,
      "median_itl_ms": 64.7,
      "p90_itl_ms": 82.7,
      "p99_itl_ms": 213.9,
      "tokens_per_second": 5190.8,
      "gpu_power_watts_total": 1863,
      "avg_output_tokens": 408,
      "under_itl_deadline": true,
      "avg_running_batch": 383.1,
      "p95_itl_ms": 97.7,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/4/384",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 4,
      "precision": "BF16",
      "max_num_seqs": 512,
      "avg_requests_in_flight": 472.8,
      "wh_per_million_output_tokens": 93.39,
      "joules_per_output_token": 0.3361983,
      "median_itl_ms": 69.1,
      "p90_itl_ms": 99.4,
      "p99_itl_ms": 335.4,
      "tokens_per_second": 5596.0,
      "gpu_power_watts_total": 1881,
      "avg_output_tokens": 402,
      "under_itl_deadline": true,
      "avg_running_batch": 472.8,
      "p95_itl_ms": 127.4,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/4/512",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "BF16",
      "max_num_seqs": 8,
      "avg_requests_in_flight": 8.0,
      "wh_per_million_output_tokens": 1464.82,
      "joules_per_output_token": 5.2733519,
      "median_itl_ms": 20.3,
      "p90_itl_ms": 20.5,
      "p99_itl_ms": 59.3,
      "tokens_per_second": 379.2,
      "gpu_power_watts_total": 2000,
      "avg_output_tokens": 406,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 20.7,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/8/8",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "BF16",
      "max_num_seqs": 16,
      "avg_requests_in_flight": 16.0,
      "wh_per_million_output_tokens": 925.3,
      "joules_per_output_token": 3.3310723,
      "median_itl_ms": 21.9,
      "p90_itl_ms": 22.3,
      "p99_itl_ms": 60.2,
      "tokens_per_second": 688.2,
      "gpu_power_watts_total": 2292,
      "avg_output_tokens": 404,
      "under_itl_deadline": true,
      "avg_running_batch": 16.0,
      "p95_itl_ms": 22.6,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/8/16",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "BF16",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 31.9,
      "wh_per_million_output_tokens": 571.95,
      "joules_per_output_token": 2.0590072,
      "median_itl_ms": 23.9,
      "p90_itl_ms": 24.4,
      "p99_itl_ms": 62.0,
      "tokens_per_second": 1243.0,
      "gpu_power_watts_total": 2559,
      "avg_output_tokens": 409,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 25.0,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/8/32",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "BF16",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.8,
      "wh_per_million_output_tokens": 350.41,
      "joules_per_output_token": 1.2614902,
      "median_itl_ms": 26.3,
      "p90_itl_ms": 27.2,
      "p99_itl_ms": 66.2,
      "tokens_per_second": 2217.6,
      "gpu_power_watts_total": 2797,
      "avg_output_tokens": 402,
      "under_itl_deadline": true,
      "avg_running_batch": 63.8,
      "p95_itl_ms": 57.5,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/8/64",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "BF16",
      "max_num_seqs": 96,
      "avg_requests_in_flight": 95.8,
      "wh_per_million_output_tokens": 267.67,
      "joules_per_output_token": 0.9636148,
      "median_itl_ms": 28.3,
      "p90_itl_ms": 29.6,
      "p99_itl_ms": 70.2,
      "tokens_per_second": 3043.3,
      "gpu_power_watts_total": 2933,
      "avg_output_tokens": 415,
      "under_itl_deadline": true,
      "avg_running_batch": 95.8,
      "p95_itl_ms": 59.4,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/8/96",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "BF16",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.6,
      "wh_per_million_output_tokens": 224.41,
      "joules_per_output_token": 0.8078852,
      "median_itl_ms": 30.3,
      "p90_itl_ms": 32.9,
      "p99_itl_ms": 78.8,
      "tokens_per_second": 3702.7,
      "gpu_power_watts_total": 2991,
      "avg_output_tokens": 397,
      "under_itl_deadline": true,
      "avg_running_batch": 127.6,
      "p95_itl_ms": 61.4,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/8/128",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "BF16",
      "max_num_seqs": 192,
      "avg_requests_in_flight": 191.5,
      "wh_per_million_output_tokens": 179.13,
      "joules_per_output_token": 0.6448575,
      "median_itl_ms": 34.3,
      "p90_itl_ms": 40.3,
      "p99_itl_ms": 105.3,
      "tokens_per_second": 4846.1,
      "gpu_power_watts_total": 3125,
      "avg_output_tokens": 405,
      "under_itl_deadline": true,
      "avg_running_batch": 191.5,
      "p95_itl_ms": 65.9,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/8/192",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "BF16",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.4,
      "wh_per_million_output_tokens": 155.93,
      "joules_per_output_token": 0.5613565,
      "median_itl_ms": 37.4,
      "p90_itl_ms": 66.2,
      "p99_itl_ms": 110.0,
      "tokens_per_second": 5695.8,
      "gpu_power_watts_total": 3197,
      "avg_output_tokens": 413,
      "under_itl_deadline": true,
      "avg_running_batch": 255.4,
      "p95_itl_ms": 69.7,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/8/256",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "BF16",
      "max_num_seqs": 384,
      "avg_requests_in_flight": 383.2,
      "wh_per_million_output_tokens": 130.59,
      "joules_per_output_token": 0.4701074,
      "median_itl_ms": 47.8,
      "p90_itl_ms": 75.3,
      "p99_itl_ms": 182.2,
      "tokens_per_second": 6782.8,
      "gpu_power_watts_total": 3189,
      "avg_output_tokens": 410,
      "under_itl_deadline": true,
      "avg_running_batch": 383.2,
      "p95_itl_ms": 78.7,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/8/384",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 8,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 8,
      "precision": "BF16",
      "max_num_seqs": 512,
      "avg_requests_in_flight": 510.9,
      "wh_per_million_output_tokens": 125.59,
      "joules_per_output_token": 0.4521216,
      "median_itl_ms": 55.8,
      "p90_itl_ms": 81.9,
      "p99_itl_ms": 237.3,
      "tokens_per_second": 6852.8,
      "gpu_power_watts_total": 3098,
      "avg_output_tokens": 403,
      "under_itl_deadline": true,
      "avg_running_batch": 510.9,
      "p95_itl_ms": 87.1,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/8/512",
      "repeat": 1
    }
  ],
  "best": {
    "H100": {
      "gpu": "H100",
      "num_gpus": 4,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 4,
      "precision": "BF16",
      "max_num_seqs": 512,
      "avg_requests_in_flight": 472.8,
      "wh_per_million_output_tokens": 93.39,
      "joules_per_output_token": 0.3361983,
      "median_itl_ms": 69.1,
      "p90_itl_ms": 99.4,
      "p99_itl_ms": 335.4,
      "tokens_per_second": 5596.0,
      "gpu_power_watts_total": 1881,
      "avg_output_tokens": 402,
      "under_itl_deadline": true,
      "avg_running_batch": 472.8,
      "p95_itl_ms": 127.4,
      "row_id": "llama-4-scout-17b-16e-instruct/H100/4/512",
      "repeat": 1
    }
  }
}