{
  "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": "qwen-3-8b",
  "name": "Qwen 3 8B",
  "model_id": "Qwen/Qwen3-8B",
  "params_b": 8.0,
  "active_params_b": 8.0,
  "architecture": "Dense Transformer",
  "gpus": [
    "B200",
    "H100"
  ],
  "configurations": [
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 8,
      "avg_requests_in_flight": 7.9,
      "wh_per_million_output_tokens": 119.55,
      "joules_per_output_token": 0.4303951,
      "median_itl_ms": 5.0,
      "p90_itl_ms": 7.4,
      "p99_itl_ms": 16.7,
      "tokens_per_second": 1501.7,
      "gpu_power_watts_total": 646,
      "avg_output_tokens": 639,
      "under_itl_deadline": true,
      "avg_running_batch": 7.9,
      "p95_itl_ms": 8.8,
      "row_id": "qwen-3-8b/B200/1/8",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 16,
      "avg_requests_in_flight": 15.9,
      "wh_per_million_output_tokens": 70.88,
      "joules_per_output_token": 0.2551701,
      "median_itl_ms": 5.8,
      "p90_itl_ms": 14.5,
      "p99_itl_ms": 29.3,
      "tokens_per_second": 2086.3,
      "gpu_power_watts_total": 532,
      "avg_output_tokens": 646,
      "under_itl_deadline": true,
      "avg_running_batch": 15.9,
      "p95_itl_ms": 19.6,
      "row_id": "qwen-3-8b/B200/1/16",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 31.9,
      "wh_per_million_output_tokens": 51.35,
      "joules_per_output_token": 0.1848479,
      "median_itl_ms": 10.3,
      "p90_itl_ms": 21.3,
      "p99_itl_ms": 34.7,
      "tokens_per_second": 2672.9,
      "gpu_power_watts_total": 494,
      "avg_output_tokens": 643,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 25.5,
      "row_id": "qwen-3-8b/B200/1/32",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.8,
      "wh_per_million_output_tokens": 34.03,
      "joules_per_output_token": 0.1225211,
      "median_itl_ms": 12.9,
      "p90_itl_ms": 23.3,
      "p99_itl_ms": 37.6,
      "tokens_per_second": 4431.7,
      "gpu_power_watts_total": 543,
      "avg_output_tokens": 643,
      "under_itl_deadline": true,
      "avg_running_batch": 63.8,
      "p95_itl_ms": 27.9,
      "row_id": "qwen-3-8b/B200/1/64",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.7,
      "wh_per_million_output_tokens": 22.78,
      "joules_per_output_token": 0.0820137,
      "median_itl_ms": 12.1,
      "p90_itl_ms": 22.4,
      "p99_itl_ms": 36.6,
      "tokens_per_second": 9118.9,
      "gpu_power_watts_total": 748,
      "avg_output_tokens": 658,
      "under_itl_deadline": true,
      "avg_running_batch": 127.7,
      "p95_itl_ms": 27.0,
      "row_id": "qwen-3-8b/B200/1/128",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.4,
      "wh_per_million_output_tokens": 20.13,
      "joules_per_output_token": 0.0724638,
      "median_itl_ms": 20.1,
      "p90_itl_ms": 32.3,
      "p99_itl_ms": 49.1,
      "tokens_per_second": 11462.6,
      "gpu_power_watts_total": 831,
      "avg_output_tokens": 654,
      "under_itl_deadline": true,
      "avg_running_batch": 255.4,
      "p95_itl_ms": 37.0,
      "row_id": "qwen-3-8b/B200/1/256",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 512,
      "avg_requests_in_flight": 511.0,
      "wh_per_million_output_tokens": 17.18,
      "joules_per_output_token": 0.0618371,
      "median_itl_ms": 34.6,
      "p90_itl_ms": 60.1,
      "p99_itl_ms": 77.7,
      "tokens_per_second": 12163.7,
      "gpu_power_watts_total": 752,
      "avg_output_tokens": 655,
      "under_itl_deadline": true,
      "avg_running_batch": 511.0,
      "p95_itl_ms": 65.8,
      "row_id": "qwen-3-8b/B200/1/512",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 1024,
      "avg_requests_in_flight": 1022.4,
      "wh_per_million_output_tokens": 17.0,
      "joules_per_output_token": 0.0612144,
      "median_itl_ms": 79.6,
      "p90_itl_ms": 110.1,
      "p99_itl_ms": 157.1,
      "tokens_per_second": 11808.6,
      "gpu_power_watts_total": 723,
      "avg_output_tokens": 641,
      "under_itl_deadline": true,
      "avg_running_batch": 1022.4,
      "p95_itl_ms": 122.4,
      "row_id": "qwen-3-8b/B200/1/1024",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 1536,
      "avg_requests_in_flight": 1426.7,
      "wh_per_million_output_tokens": 14.74,
      "joules_per_output_token": 0.0530581,
      "median_itl_ms": 91.2,
      "p90_itl_ms": 113.7,
      "p99_itl_ms": 169.2,
      "tokens_per_second": 15099.6,
      "gpu_power_watts_total": 801,
      "avg_output_tokens": 649,
      "under_itl_deadline": true,
      "avg_running_batch": 1426.7,
      "p95_itl_ms": 121.0,
      "row_id": "qwen-3-8b/B200/1/1536",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 8,
      "avg_requests_in_flight": 8.0,
      "wh_per_million_output_tokens": 123.34,
      "joules_per_output_token": 0.444018,
      "median_itl_ms": 8.2,
      "p90_itl_ms": 8.7,
      "p99_itl_ms": 13.1,
      "tokens_per_second": 961.7,
      "gpu_power_watts_total": 427,
      "avg_output_tokens": 647,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 8.8,
      "row_id": "qwen-3-8b/H100/1/8",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 16,
      "avg_requests_in_flight": 16.0,
      "wh_per_million_output_tokens": 71.85,
      "joules_per_output_token": 0.258662,
      "median_itl_ms": 8.9,
      "p90_itl_ms": 9.8,
      "p99_itl_ms": 15.2,
      "tokens_per_second": 1744.9,
      "gpu_power_watts_total": 451,
      "avg_output_tokens": 660,
      "under_itl_deadline": true,
      "avg_running_batch": 16.0,
      "p95_itl_ms": 10.7,
      "row_id": "qwen-3-8b/H100/1/16",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 31.9,
      "wh_per_million_output_tokens": 45.36,
      "joules_per_output_token": 0.1632822,
      "median_itl_ms": 10.6,
      "p90_itl_ms": 11.8,
      "p99_itl_ms": 17.4,
      "tokens_per_second": 2927.9,
      "gpu_power_watts_total": 478,
      "avg_output_tokens": 647,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 12.6,
      "row_id": "qwen-3-8b/H100/1/32",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.9,
      "wh_per_million_output_tokens": 33.12,
      "joules_per_output_token": 0.1192193,
      "median_itl_ms": 14.0,
      "p90_itl_ms": 16.5,
      "p99_itl_ms": 26.7,
      "tokens_per_second": 4314.7,
      "gpu_power_watts_total": 514,
      "avg_output_tokens": 645,
      "under_itl_deadline": true,
      "avg_running_batch": 63.9,
      "p95_itl_ms": 17.7,
      "row_id": "qwen-3-8b/H100/1/64",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 96,
      "avg_requests_in_flight": 95.8,
      "wh_per_million_output_tokens": 28.75,
      "joules_per_output_token": 0.1034956,
      "median_itl_ms": 17.8,
      "p90_itl_ms": 20.2,
      "p99_itl_ms": 32.4,
      "tokens_per_second": 5254.7,
      "gpu_power_watts_total": 544,
      "avg_output_tokens": 642,
      "under_itl_deadline": true,
      "avg_running_batch": 95.8,
      "p95_itl_ms": 21.8,
      "row_id": "qwen-3-8b/H100/1/96",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.8,
      "wh_per_million_output_tokens": 26.84,
      "joules_per_output_token": 0.0966235,
      "median_itl_ms": 20.7,
      "p90_itl_ms": 23.5,
      "p99_itl_ms": 40.7,
      "tokens_per_second": 5930.3,
      "gpu_power_watts_total": 573,
      "avg_output_tokens": 641,
      "under_itl_deadline": true,
      "avg_running_batch": 127.8,
      "p95_itl_ms": 26.7,
      "row_id": "qwen-3-8b/H100/1/128",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 192,
      "avg_requests_in_flight": 191.7,
      "wh_per_million_output_tokens": 24.05,
      "joules_per_output_token": 0.0865759,
      "median_itl_ms": 26.3,
      "p90_itl_ms": 30.8,
      "p99_itl_ms": 56.3,
      "tokens_per_second": 6932.1,
      "gpu_power_watts_total": 600,
      "avg_output_tokens": 659,
      "under_itl_deadline": true,
      "avg_running_batch": 191.7,
      "p95_itl_ms": 35.3,
      "row_id": "qwen-3-8b/H100/1/192",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.5,
      "wh_per_million_output_tokens": 22.77,
      "joules_per_output_token": 0.0819822,
      "median_itl_ms": 31.4,
      "p90_itl_ms": 36.4,
      "p99_itl_ms": 72.2,
      "tokens_per_second": 7718.8,
      "gpu_power_watts_total": 633,
      "avg_output_tokens": 654,
      "under_itl_deadline": true,
      "avg_running_batch": 255.5,
      "p95_itl_ms": 42.3,
      "row_id": "qwen-3-8b/H100/1/256",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 384,
      "avg_requests_in_flight": 383.0,
      "wh_per_million_output_tokens": 20.91,
      "joules_per_output_token": 0.0752838,
      "median_itl_ms": 42.7,
      "p90_itl_ms": 50.7,
      "p99_itl_ms": 89.2,
      "tokens_per_second": 8257.5,
      "gpu_power_watts_total": 622,
      "avg_output_tokens": 648,
      "under_itl_deadline": true,
      "avg_running_batch": 383.0,
      "p95_itl_ms": 59.3,
      "row_id": "qwen-3-8b/H100/1/384",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 512,
      "avg_requests_in_flight": 474.4,
      "wh_per_million_output_tokens": 20.04,
      "joules_per_output_token": 0.0721292,
      "median_itl_ms": 47.2,
      "p90_itl_ms": 63.7,
      "p99_itl_ms": 142.0,
      "tokens_per_second": 8534.6,
      "gpu_power_watts_total": 616,
      "avg_output_tokens": 645,
      "under_itl_deadline": true,
      "avg_running_batch": 474.4,
      "p95_itl_ms": 71.9,
      "row_id": "qwen-3-8b/H100/1/512",
      "repeat": 1
    }
  ],
  "best": {
    "B200": {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 1536,
      "avg_requests_in_flight": 1426.7,
      "wh_per_million_output_tokens": 14.74,
      "joules_per_output_token": 0.0530581,
      "median_itl_ms": 91.2,
      "p90_itl_ms": 113.7,
      "p99_itl_ms": 169.2,
      "tokens_per_second": 15099.6,
      "gpu_power_watts_total": 801,
      "avg_output_tokens": 649,
      "under_itl_deadline": true,
      "avg_running_batch": 1426.7,
      "p95_itl_ms": 121.0,
      "row_id": "qwen-3-8b/B200/1/1536",
      "repeat": 1
    },
    "H100": {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 512,
      "avg_requests_in_flight": 474.4,
      "wh_per_million_output_tokens": 20.04,
      "joules_per_output_token": 0.0721292,
      "median_itl_ms": 47.2,
      "p90_itl_ms": 63.7,
      "p99_itl_ms": 142.0,
      "tokens_per_second": 8534.6,
      "gpu_power_watts_total": 616,
      "avg_output_tokens": 645,
      "under_itl_deadline": true,
      "avg_running_batch": 474.4,
      "p95_itl_ms": 71.9,
      "row_id": "qwen-3-8b/H100/1/512",
      "repeat": 1
    }
  }
}