All Images

Introduction to GPU Computing

Figure 1

Image 1 of 1: ‘Two panels. A GPU helps when the same operation runs on thousands of elements at once, and for dense linear algebra such as matrix multiplication and convolutions. Stick with CPU when the work is sequential or full of branches and conditionals, or when the dataset is small enough that copying it to the GPU costs more than it saves. A note advises benchmarking a small case on both before requesting a GPU, since an idle GPU still blocks it for everyone else.’

Two panels. A GPU helps when the same operation runs on thousands of elements at once, and for dense linear algebra such as matrix multiplication and convolutions. Stick with CPU when the work is sequential or full of branches and conditionals, or when the dataset is small enough that copying it to the GPU costs more than it saves. A note advises benchmarking a small case on both before requesting a GPU, since an idle GPU still blocks it for everyone else.

Sagehen HPC GPU Hardware Configuration

Figure 1

Image 1 of 1: ‘Sagehen HPC has ten GPUs in three types. Four A100 cards with 80 GB, best for the largest models and highest throughput. Four L40S cards with 48 GB, the least contended and the place to start. Two RTX PRO 6000 cards with 96 GB, the most GPU memory on the cluster. A note gives the syntax for requesting a type by name, and observes that a plain --gres=gpu:1 request gives whichever type is free first and is usually quicker to schedule.’