# 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.

Source: https://greenai.guide/how-do-i/reduce-footprint-of-ai-on-a-research-cluster/  
Updated: 2026-11-19

## In short

Start by measuring: many clusters record energy per job in the Slurm scheduler, and CodeCarbon or the Green Algorithms calculator can fill gaps. Then request only the GPUs, memory and time you need, pilot ideas at small scale, checkpoint long jobs and schedule flexible ones when the grid is cleaner. Report the energy of whole projects, not just final runs.

## Steps

1. Check whether your cluster records job energy, for example in Slurm's ConsumedEnergy field.
2. Request only the GPUs, memory and wall time your job actually needs.
3. Run small pilots before scaling an experiment up.
4. Checkpoint long jobs so failures and pauses don't waste completed work.
5. Schedule flexible jobs for cleaner grid hours where the cluster allows it.
6. Report project-level energy, including tuning and failed runs.

**Shared clusters make energy easy to ignore, because nobody sees a bill.** But every over-requested GPU and
abandoned job uses power and keeps others waiting.

## How do I see what my jobs use?

- **Slurm's `sacct`** can report a job's ConsumedEnergy where the cluster has energy accounting enabled. Ask your
  research computing team if it is switched on.
- **CodeCarbon** can run inside jobs, and its documentation covers compute nodes with no internet access.
- **Green Algorithms**, from Lannelongue, Grealey and Inouye, estimates the footprint from hardware, run time,
  memory and location, which suits past jobs.

## What makes the biggest difference

- **Right-sized requests.** Asking for four GPUs and using one leaves three reserved and idle.
- **Pilots and early stopping.** Most energy goes on exploration; prune poor ideas early. See
  [cutting the energy of hyperparameter searches](/how-do-i/cut-the-energy-of-hyperparameter-searches/).
- **Checkpoints,** so crashes and time limits don't mean starting again.
- **Reuse** of pretrained models and shared datasets instead of each person preparing their own.

A 2026 UK green AI sandpit of mathematicians, computer scientists and industry also stressed checking whether a
well-understood numerical method already solves the problem before training a model at all.

## Common mistakes

- **Treating the cluster as free.** Track GPU-hours like a budget; see
  [setting a compute budget](/how-do-i/set-a-compute-budget-for-a-research-project/).
- **Reporting only the final run** in papers.

The GREENER principles from Lannelongue and colleagues offer a framework for groups and institutions. See the
[guide for researchers](/guides/researchers/).

## Sources

- [Slurm sacct manual (ConsumedEnergy)](https://slurm.schedmd.com/sacct.html), SchedMD (2025)
- [Using CodeCarbon on SLURM](https://docs.codecarbon.io/latest/how-to/slurm/), CodeCarbon documentation (2025)
- [Green Algorithms: quantifying the carbon footprint of computation](https://www.green-algorithms.org/), Lannelongue, Grealey and Inouye, Advanced Science (2021)
- [GREENER principles for environmentally sustainable computational science](https://doi.org/10.1038/s43588-023-00461-y), Lannelongue et al., Nature Computational Science (2023)
