In short
Report hardware, GPU-hours for all runs including tuning and failures, energy in kWh with the tool used, the location and grid carbon intensity, emissions in kg CO₂e, and inference energy at a stated batch size. Say what the boundary includes and how uncertain the figures are. Measuring as you go, with a tool such as CodeCarbon, makes this a table rather than a reconstruction.
Steps
- Track energy for every run from the start of the project, not just the final one.
- Report hardware, GPU-hours, energy, location, grid carbon intensity and emissions in a standard table.
- Give inference energy per 1,000 requests at a stated batch size and sequence length.
- State the boundary, including whether data-centre overheads are counted.
- Report results against compute, such as accuracy per kWh, alongside the best score.
Report compute and energy as standard results, not as a footnote. Without them, efficiency claims can't be checked, and readers can't tell whether a small accuracy gain cost ten times the energy.
The case was made early: Strubell, Ganesh and McCallum quantified the financial and environmental cost of training several language models in 2019, and Schwartz and colleagues argued in "Green AI" for reporting efficiency alongside accuracy.
What should the table include?
| Item | Example |
|---|---|
| Hardware | GPU model and count, CPU, memory |
| Run time | GPU-hours for final runs and all tuning and failed runs |
| Energy | kWh, measured or estimated, with tool and version |
| Location | Country or grid region, and carbon intensity used |
| Emissions | kg CO₂e |
| Inference | Energy per 1,000 requests at a stated batch size and sequence length |
| Boundary | Whether cooling and data-centre overheads (PUE) are included |
If you release the model, copy the figures into the Hugging Face model card's co2_eq_emissions field.
Common mistakes
- Reporting only the winning run. The search behind it is usually most of the energy.
- No uncertainty. Power-model estimates can differ from metered values; give a range.
- Leaving out negative results. An efficiency method that didn't work saves others from repeating it.
The GREENER principles from Lannelongue and colleagues give a wider framework for groups and institutions. See the guide for researchers and how to measure AI energy.
Sources
- Energy and policy considerations for deep learning in NLP, Strubell, Ganesh and McCallum, ACL (2019)
- Green AI, Schwartz, Dodge, Smith and Etzioni, Communications of the ACM (2020)
- GREENER principles for environmentally sustainable computational science, Lannelongue et al., Nature Computational Science (2023)
- Displaying carbon emissions for your model, Hugging Face Hub documentation (2026)
Related
How do I measure the carbon footprint of training a model?
Track the energy of a training run with a tool such as CodeCarbon, multiply by local grid carbon intensity, and report hardware, time and location.
How do I set a compute budget for a research project?
Set and track a compute budget for AI research: estimate GPU-hours, pilot small, stop poor runs early, reuse models and report results against compute.
Green AI for researchers
A guide for researchers on efficient AI methods, measuring and reporting energy in papers, compute budgets, funding and where maths can help.
How to measure the energy and carbon of AI
A practical guide to metering or estimating the energy and carbon of AI work: tools, system boundaries, per-task metrics and honest uncertainty.