Start Here

GPUs offer significant performance enhancements for computationally intensive tasks. GPU cores are designed for parallel computating, making them a useful tool for training machine learning models, molecular dynamics simulations, and data mining. Unlike CPUs which excel at sequential tasks, GPUs can handle large numbers of simultaneous operations.

GPU Jobs

You can view the available GPUs on Hyak with the sinfo -s command. To view which GPUs are available on the ckpt partition, use:

sinfo -p ckpt-all -O nodehost,cpusstate,freemem,gres,gresused -S nodehost | grep -v null

GPU Jobs on Checkpoint

A GPU job can be requested from ckpt by specifing the type and number of GPUs to allocate with the tag --gpus-per-node:

salloc --partition=ckpt-all --gpus-per-node=2080ti:1 --mem=10G --time=2:00:00 

GPU Jobs on a Specific GPU Partition

If you have a GPU partition, you can start an interactive session on a GPU node by using the following command:

salloc --account=account --partition=gpu-rtx6k --gpus=1 --mem=10G --time=2:00:00
# Replace the account and partition flags to match your account and partitions.

If you are unsure if your accounts have GPU partitions, use the hyakalloc command to see all of your available resources. A detailed walkthrough for requesting a GPU job can be found HERE.

GPU Types on Hyak

You now know how to view all GPUs supported on Hyak with the sinfo -s command. Additional information about each GPU is listed below:

L40 and L40s: 48GB of GDDR6 memory per GPU card

A40: 48GB of GDDR6 memory per GPU card

2080 Ti: 11GB of GDDR6 memory per GPU card

Titan: 24GB of GDDR6 memory per GPU card

RTX6k: 48GB of GDDR6 memory per GPU card

A100: 40GB of HBM2 memory per GPU card

P100: 16GB of HBM2 memory per GPU card

The next section is aimed to provide additional context for GPUs and NVIDIA NGC containers used to train LLMs.