The Chesapeake Large Scale Analytics Conference is a closed-door conference held annually where the big issues surrounding large-scale data analytics are discussed by leading experts from across the HPC space. Although attended by many of the influential figures from across the HPC community, it is geared towards the defense and intelligence communities, which have requirements that are quite unusual compared to traditional HPC.
CLSAC 2025
I was invited to attend CLSAC as a speaker on behalf of VAST. Here is what I submitted:
Title: From Exascale to Hyperscale: Insights for the Next Generation of Large-Scale Computing
Abstract:
Recent advances in generative AI have driven the construction of supercomputing infrastructure of unprecedented scale. Where the most powerful systems of the past decade deployed tens of thousands of accelerators and consumed tens of megawatts, today’s hyperscale AI supercomputers operate at ten times that scale in compute, data, power, and cost. Their physical footprints exceed those of the largest national facilities, yet they retain the tightly coupled architecture of traditional supercomputers and bring many of the same operational challenges.
In this talk, I will share an inside perspective on what large-scale AI workloads look like in practice, from the temporal patterns of months-long exascale training jobs to the physical infrastructure that supports them. Having worked both inside a major hyperscale cloud provider and now partnering closely with the newest generation of AI-focused builders, I will offer a comparative view of how different design philosophies perform at extreme scale. Drawing on first-hand experience with production hyperscale systems, I will highlight where these infrastructures excel, where they encounter fragility, and how the lessons of traditional HPC remain essential for meeting the demands of analytic grand challenges.