Application-driven considerations for the design of a multiple-processor teraflop/s computer system

Three-dimensional, time-dependent, computational fluid dynamics simulations in complex geometries are used to outline the computational needs of advanced problems in science and engineering. We discuss how these requirements could be met by a teraflop/s multiprocessor supercomputer with integrated visualization facilities. Specifically, we have in mind a parallel/vector/superscalar machine consisting of 64 CPUs. Simplified benchmarking methods, to be run on existing supercomputers, are proposed to aid the development of such a computer and its system software. Several graphics and operating system software projects for developing the facility are outlined.