GRAPE-6: A Petaflops Prototype

Abstract

We present the outline of a research project aimed at designing and constructing a hybrid computing system that can be easily scaled up to petaflops speeds. As a first step, we envision building a prototype which will consist of three main components: a general-purpose, programmable front end, a special-purpose, fully hardwired computing engine, and a multi-purpose, reconfigurable system. The driving application will be a suite of particle-based large-scale simulations in various areas of physics. The prototype system will achieve performance in the $\sim 50 - 100$ teraflops range for a broad class of applications in this area. The combination of a hardwired petaflops-class computational engine and a front end with sustained speed on the order of 10 gigaflops can produce extremely high performance, but only for the limited class of problems in which there exists a single bottleneck with computing cost dominating the total. While the calculation for which the Grape-4 (our system's immediate predecessor) was designed is a prime example of such a problem, in many other applications the primary computational bottleneck, while still related to an inverse-square (gravitational, Coulomb, etc.) force, requires less than 99% of the computing power. Although the remainder of the CPU time is typically dominated by just one secondary bottleneck, its nature varies greatly from problem to problem. It is not cost-effective to attempt to design custom chips for each new problem that arises. FPGA-based systems can restore the balance, guaranteeing scalability from the teraflops to the petaflops domain, while still retaining significant flexibility. (abbreviated abstract)