In short
Estimate the GPU-hours your main experiments need, add a margin for exploration, and track use against that budget like money. Pilot ideas at small scale, stop poor runs early, reuse pretrained models and checkpoints, and report results against the compute they took. A budget makes trade-offs visible and stops endless tuning.
Steps
- Estimate GPU-hours for the main experiments and add a stated margin for exploration.
- Run small-scale pilots to rule ideas out before scaling up.
- Track GPU-hours and energy per experiment against the budget.
- Use early stopping for poor runs instead of large grid searches.
- Reuse pretrained models and checkpoints where they fit.
- Report results against compute, such as accuracy per kWh.
Decide in advance how much computing the project may use, and track it. Without a budget, extra runs feel free, and the energy of a small accuracy gain is never weighed against it.
How do I estimate the budget?
Work from the experiments in your plan: model size, data size and number of runs give approximate GPU-hours. The Green Algorithms calculator from Lannelongue, Grealey and Inouye turns hardware, run time, memory and location into an energy and carbon estimate, which helps you see what the budget means. Add an explicit margin for exploration rather than leaving it open-ended.
What keeps a project inside budget?
- Small pilots first. Rule out ideas on small models and data subsets.
- Efficient searches. Stop poor runs early; see cutting the energy of hyperparameter searches.
- Reuse. Start from pretrained models and saved checkpoints; see avoiding training from scratch.
- Choose where and on what you run. Patterson and colleagues found that the choice of model, data centre and processor can reduce the carbon footprint of training by up to around 100 to 1,000 times.
Common mistakes
- Budgeting only the final runs. Exploration is usually most of the compute.
- No checkpoints, so a failure means starting again.
- Reporting only best accuracy, which rewards spending more.
The GREENER principles from Lannelongue and colleagues give a framework for groups and institutions. See the guide for researchers.
Sources
- GREENER principles for environmentally sustainable computational science, Lannelongue et al., Nature Computational Science (2023)
- Green Algorithms: quantifying the carbon footprint of computation, Lannelongue, Grealey and Inouye, Advanced Science (2021)
- Carbon emissions and large neural network training, Patterson et al. (2021)
Related
How do I cut the energy of hyperparameter searches?
How to tune models with less compute: smarter search than full grids, early stopping of poor trials, small pilots, good defaults and a budget.
How do I report the energy use of my AI research?
What to put in a paper about compute and energy: hardware, GPU-hours for all runs, energy, location, carbon, inference cost and the boundary used.
How do I reduce the footprint of AI experiments on a research cluster?
Cut the energy of AI experiments on HPC: read Slurm energy accounting, request only what you use, pilot small, checkpoint and schedule flexible jobs.
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.