ml-energy/benchmark-v3

The ML.ENERGY Benchmark V3 Dataset

This dataset contains benchmark results from The ML.ENERGY Benchmark, which includes LLM and diffusion inference runs on NVIDIA H100 and B200 GPUs. You can use The ML.ENERGY Leaderboard to explore the benchmarking results.

Subsets

• llm: LLM benchmark runs with full metrics (energy, throughput, latency, etc.)
• diffusion: Diffusion model benchmark runs

Usage

You can programmatically work with the dataset using the ML.ENERGY data toolkit.

pip install mlenergy-data

from mlenergy_data.records import LLMRuns, DiffusionRuns

# Load (fast, parquet only ~few MB)
runs = LLMRuns.from_hf()

# Find the most energy-efficient model on GPQA
best = min(runs.task("gpqa"), key=lambda r: r.energy_per_token_joules)
print(f"{best.nickname}: {best.energy_per_token_joules:.3f} J/tok on {best.gpu_model}")

# Column access via .data
energies = runs.data.energy_per_token_joules  # list[float]

Filter, group, and compare across GPU generations and model architectures:

# Compare GPU generations: best energy efficiency per model on GPQA
for gpu, group in runs.task("gpqa").group_by("gpu_model").items():
    best = min(group, key=lambda r: r.energy_per_token_joules)
    print(f"{gpu}: {best.nickname} @ {best.energy_per_token_joules:.3f} J/tok, "
          f"{best.output_throughput_tokens_per_sec:.0f} tok/s")

# MoE, Dense, Hybrid: who's more energy-efficient?
for arch, group in runs.task("gpqa").gpu("B200").group_by("architecture").items():
    best = min(group, key=lambda r: r.energy_per_token_joules)
    print(f"{arch}: {best.nickname} @ {best.energy_per_token_joules:.3f} J/tok")

Bulk data methods auto-download raw files as needed:

# Power timelines, output lengths, etc.
power_tl = runs.timelines(metric="power.device_instant")

Schema

LLM Runs

Field	Type	Description
task	str	Benchmark task (gpqa, lm-arena-chat, etc.)
model_id	str	Model identifier (org/model)
nickname	str	Human-readable model name
gpu_model	str	GPU model (H100, B200)
num_gpus	int	Number of GPUs used
max_num_seqs	int	Maximum batch size
energy_per_token_joules	float	Energy per output token (J)
output_throughput_tokens_per_sec	float	Output throughput (tok/s)
avg_power_watts	float	Average GPU power (W)
mean_itl_ms	float	Mean inter-token latency (ms)
is_stable	bool	Whether the run passed stability checks

See the parquet files for the full list of fields.

Diffusion Runs

Field	Type	Description
task	str	Benchmark task (text-to-image, text-to-video)
model_id	str	Model identifier (org/model)
nickname	str	Human-readable model name
gpu_model	str	GPU model (H100, B200, etc.)
num_gpus	int	Number of GPUs used
batch_size	int	Batch size
energy_per_generation_joules	float	Energy per generation (J)
height	int	Output height in pixels
width	int	Output width in pixels
batch_latency_s	float	Batch latency (s)
avg_power_watts	float	Average GPU power (W)

Issues and Discussions

Please direct any issues, questions, or discussions to The ML.ENERGY Data Toolkit GitHub repository.

Citation

@inproceedings{mlenergy-neuripsdb25,
    title={The {ML.ENERGY Benchmark}: Toward Automated Inference Energy Measurement and Optimization},
    author={Jae-Won Chung and Jeff J. Ma and Ruofan Wu and Jiachen Liu and Oh Jun Kweon and Yuxuan Xia and Zhiyu Wu and Mosharaf Chowdhury},
    year={2025},
    booktitle={NeurIPS Datasets and Benchmarks},
}

License

Apache-2.0