From Silicon to Science: The Long Road to Production Reconfigurable Supercomputing

Abstract

Over the last several years, multiple vendors have introduced systems that integrate FPGAs, as well as other types of accelerators, into machines intended for general purpose supercomputing. However, these machines have not broadly penetrated production scientific computing at any of the world’s top supercomputing centers. With the excitement around accelerators and the numerous examples of their potential, why haven’t they achieved widespread adoption in production supercomputing? This talk will discuss several barriers to adoption based on input from people who buy supercomputers and from people who use them. The short answer is that FPGA enhanced supercomputers look very little like traditional supercomputers and the performance advantage for scientific applications is often not as compelling as advertised. This talk will attempt to map barriers to adoption to specific research challenges that must be addressed to see widespread usage of FPGAs for scientific computing. These challenges include everything from the lowest level of circuit design to the programming of applications, and point to a lot of work between the current state of the art and widespread adoption of reconfigurable computing.