Using GPUs on the CQLS infrastructure

GPU-enabled software can accelerate data analysis by orders of magnitude over CPUs. Currently only NVIDIA GPUs are available on the HPC.

CQLS provides all HPC users with the following free GPU nodes:

1. cqls-gpu1 (5 Tesla V100 32GB GPUs)
2. cqls-gpu3 (4 Tesla T4 16GB GPUs)
3. cqls-p9-1 (2 Tesla V100 16GB GPUs)
4. cqls-p9-2 (4 Tesla V100 32GB GPUs)
5. cqls-p9-3 (4 Tesla V100 16GB GPUs)
6. cqls-p9-4 (4 Tesla V100 16GB GPUs)
7. ayaya01 (8 GeForce GTX 1080 Ti 11GB GPUs)
8. ayaya02 (8 GeForce GTX 1080 Ti 11GB GPUs)

Users in groups that have purchased GPU servers for group use should see them when the hqavail --gpu --no-hide-full command is run.

Note

Software will only be accelerated when you use NVIDIA GPUs if the software calls CUDA code. GPU-accelerated software will often have a GPU option (e.g., '--gpu') that can be set. Developers also often reengineer CPU-only software to take advantage of GPUs (e.g., the reimplementation of commonly used bioinformatics software by NVIDIA Parabricks). If you are uncertain if your software is using GPUs in a job, you can use the 'watch nvidia-smi' command in an interactive session to watch for GPU usage.

To use GPUs for a job, use the -G/--gpus option in hqsub to specify the number of GPUs on a node to allocate.

NVIDIA hosts a collections of containers of AI/ML GPU-enabled software that can readily be used on the HPC. If a software you'd like to use is not already installed on the HPC and is not available as a container, please submit a software ticket to let us know.