From Roadmap to Reality: Delivering HPC Solutions in an AI Dominated World

Simon Thompson, Director of Server Portfolio and Strategy at Lenovo, discusses how AI-driven accelerator and memory trends are reshaping practical HPC system design, in a session chaired by representatives of the Juelich Supercomputing Centre and the University of Illinois at Chicago. He notes Lenovo's European manufacturing site in Hungary built the first rack of its Grace Blackwell GB200 NVL4-based platform for a German customer. Comparing H200 and B200 accelerators, he shows FP4, FP8, and INT8 throughput growing while FP64 capability shifts, and presents benchmark results on FP64 emulation: it substantially improved runtime on RTX 6000 Blackwell Server Edition GPUs but slowed the GB200 NVL4, and at aggressive settings the simulation failed to converge. His central message is that most real applications are bound by memory bandwidth, I/O, or clock rate rather than FP64 matrix performance, so sites must benchmark their own workloads and adapt codes to exploit mixed precision. Addressing steep DRAM price increases, he proposes a five-step response: analyze actual memory usage (one customer found 75% of jobs fit in one gigabyte per core), buy heterogeneous memory configurations, restructure applications to avoid duplicate data copies, consider compression, and test memory channel depopulation, which recent repricing now makes cost-effective for balanced workloads. He also compares GB200 nodes favorably against CPU trays on cost and closes by arguing winners will redesign stacks for more science per gigabyte.