Codesign of Data-Centric Computing for Nuclear Security Science

Abstract

Translate Article

Data-centric computing is relied upon for nuclear security science. From simulation of complex devices using a petabyte of system memory, analysis of multi-petabyte datasets, to AI training and inference of models with hundreds of billions of parameters, these workloads push the limits of what is possible on today’s most advanced technologies. In an era in which broad performance improvements at the hardware level are largely a function of increased cost and power, the need for codesign of specialized accelerations in hardware and software has never been more acute. In this article, we describe codesign of data-centric technologies in nuclear security science. These technologies leverage the trend of near-device accelerators to address two fundamental bottlenecks in our workloads: the memory wall and the data deluge. Through the design of specialized hardware accelerations coupled with software accelerations, performance improvements of 5× to 1,000× have been achieved across a broad set of workloads.