{
  "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": "nvidia-nemotron-nano-9b-v2",
  "name": "NVIDIA Nemotron Nano 9B V2",
  "model_id": "nvidia/NVIDIA-Nemotron-Nano-9B-v2",
  "params_b": 9.0,
  "active_params_b": 9.0,
  "architecture": "Mamba-Transformer Hybrid",
  "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": 143.39,
      "joules_per_output_token": 0.5161877,
      "median_itl_ms": 5.9,
      "p90_itl_ms": 8.1,
      "p99_itl_ms": 14.1,
      "tokens_per_second": 1233.7,
      "gpu_power_watts_total": 637,
      "avg_output_tokens": 1079,
      "under_itl_deadline": true,
      "avg_running_batch": 7.9,
      "p95_itl_ms": 9.8,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 91.21,
      "joules_per_output_token": 0.3283478,
      "median_itl_ms": 7.1,
      "p90_itl_ms": 14.1,
      "p99_itl_ms": 45.9,
      "tokens_per_second": 1803.7,
      "gpu_power_watts_total": 592,
      "avg_output_tokens": 1051,
      "under_itl_deadline": true,
      "avg_running_batch": 15.9,
      "p95_itl_ms": 17.2,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 67.2,
      "joules_per_output_token": 0.2419342,
      "median_itl_ms": 10.7,
      "p90_itl_ms": 27.6,
      "p99_itl_ms": 58.5,
      "tokens_per_second": 2259.3,
      "gpu_power_watts_total": 547,
      "avg_output_tokens": 1019,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 37.3,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 53.15,
      "joules_per_output_token": 0.1913516,
      "median_itl_ms": 16.0,
      "p90_itl_ms": 33.9,
      "p99_itl_ms": 68.2,
      "tokens_per_second": 3010.1,
      "gpu_power_watts_total": 576,
      "avg_output_tokens": 1095,
      "under_itl_deadline": true,
      "avg_running_batch": 63.8,
      "p95_itl_ms": 47.8,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 44.14,
      "joules_per_output_token": 0.1589146,
      "median_itl_ms": 24.6,
      "p90_itl_ms": 52.8,
      "p99_itl_ms": 85.1,
      "tokens_per_second": 4038.0,
      "gpu_power_watts_total": 642,
      "avg_output_tokens": 1115,
      "under_itl_deadline": true,
      "avg_running_batch": 127.8,
      "p95_itl_ms": 65.9,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 255.3,
      "wh_per_million_output_tokens": 40.94,
      "joules_per_output_token": 0.1473681,
      "median_itl_ms": 40.7,
      "p90_itl_ms": 70.7,
      "p99_itl_ms": 102.1,
      "tokens_per_second": 5425.4,
      "gpu_power_watts_total": 800,
      "avg_output_tokens": 1242,
      "under_itl_deadline": true,
      "avg_running_batch": 255.3,
      "p95_itl_ms": 81.1,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 7.9,
      "wh_per_million_output_tokens": 162.73,
      "joules_per_output_token": 0.5858354,
      "median_itl_ms": 9.1,
      "p90_itl_ms": 9.3,
      "p99_itl_ms": 49.6,
      "tokens_per_second": 797.5,
      "gpu_power_watts_total": 467,
      "avg_output_tokens": 506,
      "under_itl_deadline": true,
      "avg_running_batch": 7.9,
      "p95_itl_ms": 9.3,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 15.8,
      "wh_per_million_output_tokens": 98.37,
      "joules_per_output_token": 0.3541201,
      "median_itl_ms": 10.0,
      "p90_itl_ms": 10.1,
      "p99_itl_ms": 52.4,
      "tokens_per_second": 1398.5,
      "gpu_power_watts_total": 495,
      "avg_output_tokens": 504,
      "under_itl_deadline": true,
      "avg_running_batch": 15.8,
      "p95_itl_ms": 10.5,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 31.8,
      "wh_per_million_output_tokens": 70.41,
      "joules_per_output_token": 0.2534939,
      "median_itl_ms": 12.2,
      "p90_itl_ms": 13.3,
      "p99_itl_ms": 56.6,
      "tokens_per_second": 2150.1,
      "gpu_power_watts_total": 545,
      "avg_output_tokens": 502,
      "under_itl_deadline": true,
      "avg_running_batch": 31.8,
      "p95_itl_ms": 47.2,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 63.6,
      "wh_per_million_output_tokens": 57.58,
      "joules_per_output_token": 0.2073048,
      "median_itl_ms": 16.8,
      "p90_itl_ms": 50.7,
      "p99_itl_ms": 68.7,
      "tokens_per_second": 2979.0,
      "gpu_power_watts_total": 618,
      "avg_output_tokens": 503,
      "under_itl_deadline": true,
      "avg_running_batch": 63.6,
      "p95_itl_ms": 59.0,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 95.6,
      "wh_per_million_output_tokens": 51.82,
      "joules_per_output_token": 0.1865506,
      "median_itl_ms": 21.0,
      "p90_itl_ms": 57.6,
      "p99_itl_ms": 91.6,
      "tokens_per_second": 3439.4,
      "gpu_power_watts_total": 642,
      "avg_output_tokens": 501,
      "under_itl_deadline": true,
      "avg_running_batch": 95.6,
      "p95_itl_ms": 64.9,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 127.5,
      "wh_per_million_output_tokens": 49.21,
      "joules_per_output_token": 0.1771556,
      "median_itl_ms": 25.9,
      "p90_itl_ms": 63.8,
      "p99_itl_ms": 140.6,
      "tokens_per_second": 3742.6,
      "gpu_power_watts_total": 663,
      "avg_output_tokens": 503,
      "under_itl_deadline": true,
      "avg_running_batch": 127.5,
      "p95_itl_ms": 70.4,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 191.3,
      "wh_per_million_output_tokens": 45.74,
      "joules_per_output_token": 0.1646485,
      "median_itl_ms": 34.8,
      "p90_itl_ms": 76.8,
      "p99_itl_ms": 170.1,
      "tokens_per_second": 4198.9,
      "gpu_power_watts_total": 691,
      "avg_output_tokens": 509,
      "under_itl_deadline": true,
      "avg_running_batch": 191.3,
      "p95_itl_ms": 79.3,
      "row_id": "nvidia-nemotron-nano-9b-v2/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": 255.3,
      "wh_per_million_output_tokens": 43.11,
      "joules_per_output_token": 0.1552074,
      "median_itl_ms": 43.4,
      "p90_itl_ms": 86.5,
      "p99_itl_ms": 242.5,
      "tokens_per_second": 4464.9,
      "gpu_power_watts_total": 693,
      "avg_output_tokens": 510,
      "under_itl_deadline": true,
      "avg_running_batch": 255.3,
      "p95_itl_ms": 97.6,
      "row_id": "nvidia-nemotron-nano-9b-v2/H100/1/256",
      "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": 255.3,
      "wh_per_million_output_tokens": 40.94,
      "joules_per_output_token": 0.1473681,
      "median_itl_ms": 40.7,
      "p90_itl_ms": 70.7,
      "p99_itl_ms": 102.1,
      "tokens_per_second": 5425.4,
      "gpu_power_watts_total": 800,
      "avg_output_tokens": 1242,
      "under_itl_deadline": true,
      "avg_running_batch": 255.3,
      "p95_itl_ms": 81.1,
      "row_id": "nvidia-nemotron-nano-9b-v2/B200/1/256",
      "repeat": 1
    },
    "H100": {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.3,
      "wh_per_million_output_tokens": 43.11,
      "joules_per_output_token": 0.1552074,
      "median_itl_ms": 43.4,
      "p90_itl_ms": 86.5,
      "p99_itl_ms": 242.5,
      "tokens_per_second": 4464.9,
      "gpu_power_watts_total": 693,
      "avg_output_tokens": 510,
      "under_itl_deadline": true,
      "avg_running_batch": 255.3,
      "p95_itl_ms": 97.6,
      "row_id": "nvidia-nemotron-nano-9b-v2/H100/1/256",
      "repeat": 1
    }
  }
}