TACC's Dan Stanzione on AI, Power, and the Future of Supercomputing

Podcast Timeline: Dan Stanzione (TACC) & Don Schulte00:00–02:07Introduction by Nicole; guests Dan Stanzione (Executive Director, TACC) and Don Schulte (VAST Data).02:08–03:51Reflections on TACC’s history, reputation for innovation, and pioneering adoption of new technologies.03:52–05:57Discussing dramatic shifts in HPC due to increased emphasis on power consumption, driven by the end of Dennard scaling.05:58–08:37Recent explosion of AI workload demands; increased costs and shortages (GPUs, skilled personnel, power infrastructure).08:38–12:53Speculation on future HPC developments: potential impacts of photonics, quantum computing, carbon-free energy sources, and changes in AI scaling strategies.12:54–18:20Dan emphasizes the importance of foundational HPC research historically done at national labs and universities, highlighting that current AI and infrastructure innovations rely heavily on these early HPC breakthroughs.18:21–21:49Introduction of Horizon, TACC’s upcoming NSF-funded supercomputer, replacing the Frontera system, focusing on scientific throughput, GPU optimization, and extensive solid-state storage.21:50–27:57Detailed discussion on the NSF’s Leadership Class Computing Facility (LCCF) award that supports Horizon, emphasizing scientific outcomes over raw computing power.Horizon system designed specifically for real-time data assimilation, persistent interactive services, and complex scientific workflows, enabling significant improvements in science productivity.27:58–30:36Shift from batch-oriented computing to interactive, real-time workflows and persistent data management.Importance of new data platforms (like VAST) providing consistent, high-performance data access across diverse computing tasks.30:37–34:47Stanzione emphasizes new data access patterns: smaller, random, constant I/O operations, challenging traditional HPC storage assumptions. Highlights VAST’s platform role in addressing these new storage needs effectively.34:48–36:33Closing remarks on the dramatic evolution in HPC data management over the past decade, noting fundamental shifts that were not anticipated even ten years ago.