GPU Partition Changes

HCC offers 4 different types of GPU units on the Crane supercomputing cluster for individuals interested in compiling and executing CUDA or OpenACC code.
HCC offers 4 different types of GPU units on the Crane supercomputing cluster for individuals interested in compiling and executing CUDA or OpenACC code.

Holland has changed the configuration of the GPU partition available on the Crane cluster. As an alternative to requesting specific types of GPU units, users now have the option to submit jobs to any available node with a graphics processing unit. Additionally, those wishing to continue constraining their jobs to particular unit types have the ability to specify which type(s) they wish to utilize.

To run jobs on these nodes, the following directives will need to be included in srun commands (for interactive jobs):
--partition=gpu --gres=gpu

Or the inclusion of the lines:
#SBATCH --partition=gpu
#SBATCH --gres=gpu
in submit scripts (for batch jobs).

The job can be further constrained to a specific type of unit by the addition of the directive:
--constraint=GPU_type
where GPU_type can be one of the following: gpu_m2070, gpu_k20, gpu_k40, gpu_p100.

In a submit script, this directive would look like:
#SBATCH --constraint=GPU_type

For more information, including example submit scripts and interactive job commands, see the documentation.