GPUs on M3
M3 provides users access to a variety of modern NVIDIA GPUs. All M3 users will have access to GPUs in the gpu and desktop partitions (desktop is only to be used by Strudel jobs).
You may notice GPUs in other partitions, but these are most likely restricted. See the restricted partitions table to understand if you can request access to a given partition.
How many GPUs does M3 have?
M3's GPU makeup is constantly evolving as we acquire new hardware and retire old hardware. The below table is accurate as of the time of writing this.
| Partition | GPU type | Number of this GPU |
|---|---|---|
gpu | A40 | 12 |
gpu | A100 | 32 |
gpu | T4 | 8 |
desktop | A40 | 44 |
desktop | P4 | 60 |
desktop | T4 | 54 |
How do I access a GPU on M3?
On Strudel
If you are using Strudel, you should see a dropdown for the desired app (e.g. Desktop or Jupyter Lab) which lets you select "No GPUs" or any single GPU type available in the desktop partition.
On the command line
See Specifying resources in Slurm for more details on sbatch commands.
To request a single GPU from the gpu partition:
sbatch --partition=gpu --gres=gpu:1 ...
To request a single A40 GPU from the gpu partition:
sbatch --partition=gpu --gres=gpu:A40:1 ...
To request two A40 GPUs from the gpu partition:
sbatch --partition=gpu --gres=gpu:A40:2 ...
Troubleshooting
Why don't I seem to have a GPU even though I'm in the gpu partition?
This could be because you forgot to specify --gres=gpu... in your sbatch job specifications. Use the nvidia-smi command to confirm that you have access to a GPU.