Isamabard AI: A National Exascale AI Research Resource Inside a Containerised Data Center
Isambard AI supercomputing platform has been designed for sustainability and performance since AI applications demand unprecedented levels of computing power. Several innovative capabilities have been brought together to deliver, in a record time, a national AI research resource (AI RR) at the University of Bristol, UK. This talk provides an overview of the underlying technologies for a highly sustainable solution, which is based on a HPE modular data centre (MDC) product called Performance Optimised Datacentre (POD), Nvidia ARM based Grace-Hopper superchip-based nodes, HPE direct liquid cooled EX series cabinets with SlingShot interconnet, and VAST software defined storage system. The resulting solution was not only put together from scratch in a matter of a few months but also is delivering high performance and power efficiencies for AI and ML benchmarks. Furthermore, Isambard has been designed from ground up to be a federated, cloud-native supercomputing ecosystem to lower access barriers to a national leadership-class AI RR for users from diverse backgrounds and communities.