GPU class jobs can be requested by selecting from one of the below configurations

GPU typeNumber of gpus nncpus/gpu Xmem/GB YMax walltime/hr
 K80, P100, RTX6000 1,2,4,8 4n  24n 24
 P1000  1  1-8  96  72
Summary of the table's contents

Where n is a multiplier matching the number of GPUs requested. The additional options ngpus= and gpu_type= must be added to the PBS selection.

For example, to specify a job using 4 RTX6000 GPUs, the selection must be

#PBS -lselect=1:ncpus=16:mem=96gb:ngpus=4:gpu_type=RTX6000

Within the context of the running job, the shell environment variable CUDA_VISIBLE_DEVICES will be set with indices of the allocated GPUs. Jobs must respect this setting, or they will interfere with other jobs co-located on the execution node

The details of the different GPU types are:

GPU type

Single precision /TFLOPS

Double precision/TFLOPS

Memory bandwidth/GB/s
Recommended use
P1000 1.8 <<1 4 80

Interactive use via Jupyter

Ensemble tasks: ML inference, molecular dynamics, image processing, etc

K80 5.6 2.9 12 240 Numerical simulation
P100 8.0 4.0 16 730 Numerical simulation
RTX6000 16.3 <1 24 670 ML training