Using GPUs on Firebird

Firebird offers multiple nodes with NVIDIA GPUs. Typically, these must be requested via Slurm before they can be used. There are different methods to request one or more GPUs depending on whether you are submitting a batch script, working interactively on the command line, or creating a virtual desktop session.

List of GPUs

The following table lists which specific types of GPUs are available, some of their characteristics, and which nodes they are in. Examples of how to use this information to request a specific GPU are below.

Node	GPU Type and Number	Memory
gpu01	rtx2080ti (4)	11 GB
gpu02	quadro_8000 (3)	48 GB
gpu03	rtx2080ti (4)	11 GB
gpu04	l40s (4)	48 GB
gpu05	h200 (2)	141 GB
gpu06	rtx6000 (2)	96 GB

Requesting a GPU in a Batch Script or Interactively

The easiest way to request a GPU through Slurm is with the --gpus-per-nodedirective.

If you do not have a preference which specific GPU you want:

Within a job submission script:

#SBATCH --gpus-per-node=1

As part of an salloccommand to create an interactive terminal session:

--gpus-per-node=1

If you want to request a specific model of GPU, add the type from the table above, for example:

--gpus-per-node=h200:1

In addition to requesting the GPU, it is necessary to load the CUDA module. Either within a job submission script, or from the command line:

module load cuda

This will load the latest default CUDA version. To see a list of all available CUDA versions, for example:

module keyword cuda

The following modules match your search criteria: "cuda"
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

  cuda: cuda/11.8-0, cuda/12.4-0

Requesting a GPU in a Virtual Desktop Session

As part of the form to request a virtual desktop session, there is a drop-down box to request a specific model of GPU, if desired.