In short
Measure the current system before you change anything, on the same hardware, load and inputs you will use afterwards, and repeat the runs to see the normal spread. Report energy per unit of useful work as well as in total, and only claim a saving bigger than the noise and achieved without a drop in quality.
Steps
- Choose a functional unit, such as energy per request or per document.
- Measure the current system several times on a fixed set-up and workload, including the idle baseline.
- Change one thing at a time and re-measure under the same conditions.
- Record quality or task success alongside energy for each run.
- Claim only savings larger than the run-to-run spread, and report totals as well as per-unit figures.
A saving only counts if you measured before, measured after the same way, and the difference is bigger than the noise. Without a baseline, an energy claim is a guess, and it may not survive scrutiny.
What makes a fair comparison?
- Same set-up: hardware, region, software versions and settings.
- Same load: the same requests, documents or training data, at the same volume.
- Several runs: readings vary with temperature, background processes and frequency scaling. Report the mean and range.
- Energy, not just speed: the GreenCode Constitution work found, on the llama.cpp inference engine, a 4.7% cut in GPU energy alongside only a 3% speed gain. Time and energy can tell different stories.
- Quality held: a cheaper answer that is wrong more often is not a saving.
Which unit should the baseline use?
Use a functional unit, as in the Green Software Foundation's SCI specification: per request, per document, or per successful task. Then report the total alongside it, because energy per task can fall while total energy rises as use grows. See tracking AI energy per task.
Common mistakes
- Measuring once, before and after. One pair of readings can be noise.
- Changing several things at once, so you can't tell what helped.
- Moving the baseline, for example comparing a new small model with an old large one at a different volume.
- Counting carbon falls that come only from a cleaner grid as efficiency.
Before you publish a saving, check it against honest green claims.
Sources
- Software Carbon Intensity (SCI) specification, Green Software Foundation (2024)
- GreenCode Constitution: agent principles for optimizing code and LLM inference energy use, Kapenekakis, Kim, Ionescu, Thomsen and Albano, GreenAI workshop, ECML PKDD 2026 (2026)
Related
How do I track AI energy per task rather than in total?
Pick a functional unit such as per document or per successful task, log energy and quality against it, and report it next to the total.
How do I set targets for AI energy and emissions?
How to set AI energy and emissions targets that mean something: a measured baseline, intensity and absolute goals together, and a regular review.
How do I make honest claims about our green AI?
How to describe your AI's environmental performance without greenwashing: specific, measured, comparable claims with a baseline, a method and the limits.
Honest green claims about AI
How to avoid greenwashing about AI: what you can and can't claim, rebound effects, honest reporting, and the UK, EU and US rules on environmental claims.