Abstract

The United Kingdom’s AI Research Resource (AIRR) currently consists of two supercomputers: Dawn, with Intel GPUs, at the University of Cambridge; Isambard-AI, with NVIDIA GP Us, at the University of Bristol. Future additions to AIRR may have GPUs from other manufacturers, or may have alternative acceleration devices. For researchers aiming to run AI software that they develop themselves, it can be useful to be able to perform development and testing on a local machine, and to be able to run exactly the same software on any AIRR system, taking advantage of the GPUs available, also across multiple nodes, with data and models distributed as needed.

This talk will review device selection and processing strategies in AI workflows, will outline how different devices and strategies can be handled explicitly in PyTorch code, and will show how devices and strategies supported by PyTorch can be handled implicitly with Lightning. There will be a demonstration of how the same Lightning-based AI application can be run on a laptop, on Dawn, and on Isambard-AI.