AI energy benchmarks compared
Three public benchmarks measure AI energy in different ways. Knowing each one's boundary is what makes their numbers usable.
At a glance
| Benchmark | Run by | What it measures | Boundary |
|---|---|---|---|
| MLPerf Power | MLCommons | Energy of benchmark runs, reported as samples per joule only for results that meet the quality target | Data-center systems including compute nodes and interconnect; excludes cooling and storage nodes |
| AI Energy Score | Hugging Face and collaborators | Energy of models on a fixed set of tasks and samples, as a comparative rating | The hardware running the model under controlled conditions |
| ML.ENERGY benchmark | The ML.ENERGY initiative | Energy of generative AI inference across models and settings | GPU-level measurement |
Sources: Tschand et al., MLPerf Power; Hugging Face, announcing the AI Energy Score, 11 February 2025; ML.ENERGY benchmark paper.
How the Joule Ledger relates
- MLPerf Power's data-center boundary corresponds to JL-26 meter point M3, and JL-26 adopts its instrument rules: SPEC analyzers, PTDaemon, ranging runs and 60-second windows.
- Like MLPerf, JL-26 counts only quality-passing results in Decision Yield.
- JL-26 adds the facility (M4), the closure check, and labels and grades for every number.
Choosing one
Use a benchmark to compare published systems on a common task. Use the Joule Ledger when you need to account for your own system's energy, at your own boundaries, in a form others can check.
Trademarks
MLPerf is a trademark of MLCommons. Other names are trademarks of their owners. AAERL is not affiliated with any benchmark named here.
Questions
Can MLPerf results be compared with facility PUE?
No. MLPerf Power excludes cooling, so its figures sit at the system boundary; PUE is a facility ratio.
Sources
Keep reading
Take the next step.
MLPerf Power, the AI Energy Score and the ML.ENERGY benchmark compared: what each measures, where its boundary sits, and how the Joule Ledger relates.