Portable and Resilient by Design: A Cloud-Style Stack for HPC and AI across NVIDIA Grace and Grace Hopper Superchips

Abstract

The Bristol Centre for Supercomputing (BriCS) operates two production systems based on the NVIDIA Superchip family that together serve the UK’s national HPC and AI programmes. Isambard 3 is a 384-node Grace CPU Superchip system (55,296 Arm Neoverse V2 cores) supporting memory-bandwidth-bound simulation for the national HPC service, UKRI open-access calls, and GW4 partners. Isambard-AI is a Grace Hopper GH200 system—the UK’s AI Research Resource (AIRR)—delivering large-scale training and inference to thousands of users and over a thousand projects across government, academia, industry, and start-ups. A single coherent-memory Arm architecture thus spans two very different points in the workload spectrum. This paper presents our state of practice in operating both. Our central finding is that an open, cloud-style software and delivery stack across the two systems—open-standards software with CI/CD-based delivery, Slurm and Kubernetes orchestration, a cloud-like single front door for access, federated identity, and an emphasis on documentation and training—is what enabled successful delivery to a large and diverse user base, and, crucially, what makes the facility portable and resilient: not locked to one hardware or software vendor—a measure of technical sovereignty, with x86-to-aarch64 portability already proven—and able to keep pace with frequent AI changes through cloud-style operations, multi-tenancy, and rapid upgrades. We report selected performance results for simulation and AI/ML workloads as validation, alongside programmatic delivery metrics and energy-efficiency practice (direct liquid cooling, PUE below 1.08). We close on how this portable stack positions us for evolving Arm/Superchip hardware, including Vera-Rubin and Arm AGI, and application readiness.