Skip to content

GPU access and partitions

Overview

GPU resources provide accelerated computation for suitable workloads.

They are limited and managed separately from standard CPU resources.


Access

GPU access is not enabled by default.

You must be granted access before submitting GPU jobs.

Jobs that request GPUs without access will not start.


GPU partitions

GPU-enabled nodes are grouped into specific partitions, such as:

  • l40s
  • a100

Check available partitions:

sinfo

Requesting GPUs

Example job configuration:

#SBATCH --partition=l40s
#SBATCH --gres=gpu:l40s:1

The partition and GPU type must match.


Behaviour

GPU jobs:

  • may queue longer due to limited availability
  • require correct configuration to use GPU resources

Not all workloads benefit from GPUs.


Common issues

Jobs may not start if:

  • GPU access has not been granted
  • incorrect partition is specified

GPUs may not be used if:

  • software is not configured for GPU execution

Good practice

  • test workflows on CPU first
  • request only the GPUs required
  • confirm configuration before scaling