{
  "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": "gemma-3-27b",
  "name": "Gemma 3 27B",
  "model_id": "google/gemma-3-27b-it",
  "params_b": 27.0,
  "active_params_b": 27.0,
  "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": 8.0,
      "wh_per_million_output_tokens": 376.14,
      "joules_per_output_token": 1.3541189,
      "median_itl_ms": 13.2,
      "p90_itl_ms": 15.5,
      "p99_itl_ms": 25.7,
      "tokens_per_second": 594.2,
      "gpu_power_watts_total": 805,
      "avg_output_tokens": 718,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 17.3,
      "row_id": "gemma-3-27b/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": 197.49,
      "joules_per_output_token": 0.710946,
      "median_itl_ms": 13.5,
      "p90_itl_ms": 18.0,
      "p99_itl_ms": 35.8,
      "tokens_per_second": 1125.2,
      "gpu_power_watts_total": 800,
      "avg_output_tokens": 721,
      "under_itl_deadline": true,
      "avg_running_batch": 15.9,
      "p95_itl_ms": 21.1,
      "row_id": "gemma-3-27b/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": 129.51,
      "joules_per_output_token": 0.4662487,
      "median_itl_ms": 15.8,
      "p90_itl_ms": 23.1,
      "p99_itl_ms": 44.6,
      "tokens_per_second": 1844.0,
      "gpu_power_watts_total": 860,
      "avg_output_tokens": 717,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 28.2,
      "row_id": "gemma-3-27b/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": 81.59,
      "joules_per_output_token": 0.2937082,
      "median_itl_ms": 19.3,
      "p90_itl_ms": 28.4,
      "p99_itl_ms": 47.4,
      "tokens_per_second": 3062.1,
      "gpu_power_watts_total": 899,
      "avg_output_tokens": 712,
      "under_itl_deadline": true,
      "avg_running_batch": 63.8,
      "p95_itl_ms": 35.1,
      "row_id": "gemma-3-27b/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": 127.8,
      "wh_per_million_output_tokens": 59.92,
      "joules_per_output_token": 0.2157049,
      "median_itl_ms": 26.4,
      "p90_itl_ms": 35.6,
      "p99_itl_ms": 60.4,
      "tokens_per_second": 4438.8,
      "gpu_power_watts_total": 957,
      "avg_output_tokens": 714,
      "under_itl_deadline": true,
      "avg_running_batch": 127.8,
      "p95_itl_ms": 40.9,
      "row_id": "gemma-3-27b/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": 247.3,
      "wh_per_million_output_tokens": 49.1,
      "joules_per_output_token": 0.1767455,
      "median_itl_ms": 39.6,
      "p90_itl_ms": 54.3,
      "p99_itl_ms": 111.3,
      "tokens_per_second": 5510.4,
      "gpu_power_watts_total": 974,
      "avg_output_tokens": 717,
      "under_itl_deadline": true,
      "avg_running_batch": 247.3,
      "p95_itl_ms": 63.8,
      "row_id": "gemma-3-27b/B200/1/256",
      "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": 419.68,
      "joules_per_output_token": 1.5108348,
      "median_itl_ms": 23.4,
      "p90_itl_ms": 23.8,
      "p99_itl_ms": 24.9,
      "tokens_per_second": 337.8,
      "gpu_power_watts_total": 510,
      "avg_output_tokens": 719,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 24.0,
      "row_id": "gemma-3-27b/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": 239.83,
      "joules_per_output_token": 0.8633876,
      "median_itl_ms": 25.1,
      "p90_itl_ms": 25.9,
      "p99_itl_ms": 29.6,
      "tokens_per_second": 624.6,
      "gpu_power_watts_total": 539,
      "avg_output_tokens": 726,
      "under_itl_deadline": true,
      "avg_running_batch": 16.0,
      "p95_itl_ms": 26.2,
      "row_id": "gemma-3-27b/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": 31.7,
      "wh_per_million_output_tokens": 145.87,
      "joules_per_output_token": 0.5251489,
      "median_itl_ms": 28.3,
      "p90_itl_ms": 29.1,
      "p99_itl_ms": 47.1,
      "tokens_per_second": 1088.2,
      "gpu_power_watts_total": 571,
      "avg_output_tokens": 714,
      "under_itl_deadline": true,
      "avg_running_batch": 31.7,
      "p95_itl_ms": 29.9,
      "row_id": "gemma-3-27b/H100/1/32",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "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": 430.19,
      "joules_per_output_token": 1.5486857,
      "median_itl_ms": 15.1,
      "p90_itl_ms": 15.4,
      "p99_itl_ms": 16.8,
      "tokens_per_second": 525.0,
      "gpu_power_watts_total": 813,
      "avg_output_tokens": 711,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 15.6,
      "row_id": "gemma-3-27b/H100/2/8",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "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": 247.56,
      "joules_per_output_token": 0.8912014,
      "median_itl_ms": 16.1,
      "p90_itl_ms": 16.4,
      "p99_itl_ms": 19.8,
      "tokens_per_second": 978.3,
      "gpu_power_watts_total": 872,
      "avg_output_tokens": 707,
      "under_itl_deadline": true,
      "avg_running_batch": 16.0,
      "p95_itl_ms": 16.5,
      "row_id": "gemma-3-27b/H100/2/16",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 32.0,
      "wh_per_million_output_tokens": 154.0,
      "joules_per_output_token": 0.5544002,
      "median_itl_ms": 18.4,
      "p90_itl_ms": 18.9,
      "p99_itl_ms": 28.4,
      "tokens_per_second": 1686.9,
      "gpu_power_watts_total": 935,
      "avg_output_tokens": 710,
      "under_itl_deadline": true,
      "avg_running_batch": 32.0,
      "p95_itl_ms": 19.2,
      "row_id": "gemma-3-27b/H100/2/32",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.9,
      "wh_per_million_output_tokens": 107.78,
      "joules_per_output_token": 0.3880231,
      "median_itl_ms": 22.8,
      "p90_itl_ms": 23.7,
      "p99_itl_ms": 51.5,
      "tokens_per_second": 2653.7,
      "gpu_power_watts_total": 1030,
      "avg_output_tokens": 711,
      "under_itl_deadline": true,
      "avg_running_batch": 63.9,
      "p95_itl_ms": 25.2,
      "row_id": "gemma-3-27b/H100/2/64",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 96,
      "avg_requests_in_flight": 95.8,
      "wh_per_million_output_tokens": 91.47,
      "joules_per_output_token": 0.3293027,
      "median_itl_ms": 27.0,
      "p90_itl_ms": 28.5,
      "p99_itl_ms": 63.6,
      "tokens_per_second": 3344.1,
      "gpu_power_watts_total": 1101,
      "avg_output_tokens": 711,
      "under_itl_deadline": true,
      "avg_running_batch": 95.8,
      "p95_itl_ms": 33.0,
      "row_id": "gemma-3-27b/H100/2/96",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.8,
      "wh_per_million_output_tokens": 85.04,
      "joules_per_output_token": 0.3061552,
      "median_itl_ms": 32.0,
      "p90_itl_ms": 34.5,
      "p99_itl_ms": 90.5,
      "tokens_per_second": 3714.0,
      "gpu_power_watts_total": 1137,
      "avg_output_tokens": 705,
      "under_itl_deadline": true,
      "avg_running_batch": 127.8,
      "p95_itl_ms": 40.4,
      "row_id": "gemma-3-27b/H100/2/128",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 192,
      "avg_requests_in_flight": 187.0,
      "wh_per_million_output_tokens": 80.42,
      "joules_per_output_token": 0.2895195,
      "median_itl_ms": 39.8,
      "p90_itl_ms": 45.6,
      "p99_itl_ms": 141.2,
      "tokens_per_second": 4180.8,
      "gpu_power_watts_total": 1210,
      "avg_output_tokens": 714,
      "under_itl_deadline": true,
      "avg_running_batch": 187.0,
      "p95_itl_ms": 62.6,
      "row_id": "gemma-3-27b/H100/2/192",
      "repeat": 1
    }
  ],
  "best": {
    "B200": {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 247.3,
      "wh_per_million_output_tokens": 49.1,
      "joules_per_output_token": 0.1767455,
      "median_itl_ms": 39.6,
      "p90_itl_ms": 54.3,
      "p99_itl_ms": 111.3,
      "tokens_per_second": 5510.4,
      "gpu_power_watts_total": 974,
      "avg_output_tokens": 717,
      "under_itl_deadline": true,
      "avg_running_batch": 247.3,
      "p95_itl_ms": 63.8,
      "row_id": "gemma-3-27b/B200/1/256",
      "repeat": 1
    },
    "H100": {
      "gpu": "H100",
      "num_gpus": 2,
      "tensor_parallel": 2,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 192,
      "avg_requests_in_flight": 187.0,
      "wh_per_million_output_tokens": 80.42,
      "joules_per_output_token": 0.2895195,
      "median_itl_ms": 39.8,
      "p90_itl_ms": 45.6,
      "p99_itl_ms": 141.2,
      "tokens_per_second": 4180.8,
      "gpu_power_watts_total": 1210,
      "avg_output_tokens": 714,
      "under_itl_deadline": true,
      "avg_running_batch": 187.0,
      "p95_itl_ms": 62.6,
      "row_id": "gemma-3-27b/H100/2/192",
      "repeat": 1
    }
  }
}