Panel Discussion + Closing

This closing panel of a vendor-focused session on HPC and AI infrastructure brings together representatives from NVIDIA, Lenovo, Bull/Eviden, and Sugon, moderated by session chairs from the Juelich Supercomputing Centre and the University of Illinois at Chicago. An audience question about running FP64-heavy weather and climate simulations in an era of FP4/FP8-oriented accelerators draws responses that most such applications are bound by memory bandwidth rather than flops, that architecture decisions are informed by co-analysis with application developers, that NVIDIA GPU FP64 hardware remains IEEE compliant with library-level safeguards around reduced-precision methods, and that CPU product lines continue alongside GPUs. The Bull panelist describes the BXI interconnect: a hardware implementation of the Portals library with full offload, automated congestion control, dynamic rerouting around hotspots, and UEC compliance planned for the next generation. Asked to predict how scientists will use HPC by 2030, panelists foresee agentic AI workflows that read literature, configure and run jobs, platforms letting users build domain-specific agents, greater hardware heterogeneity, maturing quantum computing integrated via offloaded kernels, and a bold claim that most supercomputer cycles will be launched through natural language, while stressing scientists must retain critical thinking. Discussion of AI agents on login nodes weighs security and multi-tenancy concerns against the view that agents are no riskier than students. Closing advice urges centers to prepare for liquid cooling, power density, memory price volatility, and deep workload understanding.