Closing Keynote: HPC in Transition

In this closing keynote, Jack Dongarra (University of Tennessee, Oak Ridge National Laboratory) surveys how generative AI is reshaping high-performance computing. He notes AI capital investment at scales that challenge nations, citing a Memphis system planned with one million NVIDIA Blackwell GPUs, and trillion-dollar market capitalizations that dwarf traditional HPC vendors. He argues HPC and AI are converging into hybrid workflows where AI surrogates rapidly explore parameter spaces and physics-based simulation verifies results, citing weather forecasting at ECMWF and protein structure prediction. Dongarra emphasizes that energy and data movement, not floating-point operations, are the scarce resources: top systems reach under 1% of theoretical peak on HPCG, and he proposes metrics like joules per trusted solution over FLOPS, criticizing aging benchmarks that reward incremental change. He contrasts commodity-driven procurement with genuine co-design, holding up the CPU-only LionShine system and Fugaku as exceptions, and asks whether GPUs are truly necessary. Tracing NVIDIA generations from Hopper to Blackwell to Rubin, he shows 64-bit performance stagnating or declining while 16-bit tensor throughput grows, and demonstrates mixed-precision iterative refinement that solves dense linear systems roughly eight times faster with eight times less energy at full accuracy. In the Q&A he discusses Ozaki-style FP64 emulation, the need for more experimental architectures, photonics as a long-term path to energy efficiency, and the need for more numerical analysis expertise.