Supercomputing after Device Scaling Stops: Limits, Tradeoffs, and What Comes Next

Abstract

Moore’s Law and Dennard scaling no longer give us an easy path to “more-of-the-same, but better” supercomputers. As a result, the supercomputing community faces a new era increasingly defined by energy budgets, data movement costs, and architectural tradeoffs. The explosion of very large systems optimized for AI is also reshaping the HPC landscape. In this talk, I’ll explore the challenges and opportunities that lie ahead for supercomputing in a post-scaling world: how energy density will shape the physical limits of future systems, how memory hierarchies must evolve under bandwidth and energy constraints, why data placement is becoming a first-class citizen in software design, and why adapting code to utilize the low-precision datapaths of AI accelerators may prove worth the effort.