Deploying Next-Generation HPC Systems with NVIDIA Superchips: Early Experiences, Challenges, and Best Practices

The introduction of NVIDIA's Grace Hopper and Grace Blackwell Superchips marks a significant architectural shift in GPU architecture as HPC centers seek to both balance traditional simulation workloads with AI and data-intensive applications. With early adopters now deploying these CPU-GPU accelerated platforms into production environments, the community is navigating a distinct set of opportunities and challenges spanning system design, energy efficiency, workflow portability, application readiness and user education.

This BoF, organized by the NVidia Supercomputing Users Group (NVSUG), aims to bring together HPC practitioners and system architects who have deployed or plan to deploy systems based on Grace Hopper, Grace Blackwell, or Vera Rubin architectures. Computer scientists will also be invited to contribute with their users’ experiences and perspectives. The session will focus on sharing practical experiences from the field, including architectural considerations, procurement lessons learned, integration with existing clusters, and early insights into application behavior and performance. The organizers will begin with a series of short presentations describing current experiences and plans. We will cover our efforts to deploy ALPS (CSCS), Doudna (NERSC), Horizon (TACC), Isambard-AI (Bristol), Miyabi (JCAHPC), and GENCI-IDRIS (Dalia) for the first 20 minutes. The remaining time (40 minutes) will be dedicated for open discussion.

We invite attendees from HPC centers of all sizes - from national laboratories and supercomputing facilities to academic and industry-focused research groups to share their experiences and participate in the discussions. The goal of this BoF is to foster an open dialogue between HPC centers using or considering NVIDIA GPU technologies, sharing best practices and lessons-learned. Attendees will gain insights into the latest NVIDIA Superchip deployments, practical best practices, and new opportunities for joint collaboration.

The NVidia Supercomputing User Group (NVSUG) is an independent, non-profit organization of supercomputing centers powered by NVIDIA technologies, established to provide a platform for collaboration and information exchange. It is a community-driven group, covering all topics related to High Performance Computing (HPC), Artificial Intelligence (AI) and Quantum Computing. The NVSUG was initiated during ISC24 and it was followed by a formal kick-off at SC24 with 35 Supercomputing Sites and 65+ attendees. New participant sites joined at ISC25 and SC25 meetings, each event supporting 60+ attendees. NVSUG's first BoF has been accepted at SC Asia 2025 “How to keep up with AI’s rapid transformation?” and as a community group, NVSUG is planning to engage with the wider HPC community through additional BoFs on major conferences like ISC26, discussing topics raised as of high interest.

---

Organizers:

• John Cazes
• Colin McMurtie
• Taisuke Boku
• Crhistelle Piechurski
• Sadaf Alam