Available GPUs

GPUs are available in three locations:

  • All login nodes are equipped with a small GPU to allow compiling and testing GPU applications.
  • The UV300 has 4 x nVidia Tesla P100 GPUs.
  • One batch partition node has 2 x nVidia Tesla A100 GPUs.

Requesting GPUs

Job requests for GPUs need to specify at least the following three Slurm options:

  • An account with access to the nih_s10 partition, usually one of:
    • --account=your_PI_SUNetID
    • --account=your_Project_ID

  • The partition:

    • --partition=nih_s10 OR --partition=batch

  • For N gpus (1 <= N <= 4):

    • --gres=gpu:N

CUDA

There are several versions of CUDA available to use, for a complete list run:

  • module avail cuda

There are also several CUDA enabled python virtual environments, to see a list run:

  • module load anaconda; conda env list | grep cuda