The ML.ENERGY Benchmark produces a rich dataset: energy, latency, throughput, and power traces across models, GPUs, and tasks. To make this data easy to use, we built The ML.ENERGY Data Toolkit. Like always, as open-source as it can be.

The toolkit lets you programmatically load and filter benchmark runs with typed, immutable collections, extract bulk data as DataFrames, and fit models like logistic power curves and latency distributions.

We use it to compile data for The ML.ENERGY Leaderboard and run analyses for our blog. More importantly, it is enabling external use cases that require realistic LLM inference power & energy data, like OpenG2G.

OpenG2G is an AI datacenter-grid coordination simulation framework, and it uses the ML.ENERGY Benchmark data and the toolkit feed prooduction-grade GPU power traces into simulation. We'll share more on the OpenG2G use case in an upcoming update.

If you want to get started, check out the documentation for a usage guide and API reference. Raw data is hosted on Hugging Face Hub.