The Performance of the Intel TFLOPS Supercomputer

The purpose of building a supercomputer is to provide superior performance on real applications. In this paper, we describe the performance of the Intel TFLOPS Supercomputer starting at the lowest level with a detailed investigation of the Pentium® Pro processor and the supporting memory subsystem. We follow this with a description of the benchmarks used to track the performance of the machine over its development life cycle, which culminated in the first MP LINPACK run to exceed a rate of one trillion floating point operations per second (TFLOPS). Our analysis applies not only to the TFLOPS supercomputer, but also to servers and workstations based on the Intel 32-bit architecture. We conclude with a discussion of the machine's performance on a production application.