Abstract

It is now recognized that technological improvements are slowing down and that to obtain high performance in computer systems we must have recourse to parallelism. The most powerful supercomputers of today (e.g., the Cray XMP) consist of a small number of tightly coupled multifunctional unit (scalar and vector) processors sharing a large main memory. These systems are intended for general scientific computations. Systems for more limited applications (such as the MPP for image processing) have been built with thousands of simple (1-bit ALU) mesh-connected processors. To meet the requirements of future computing utilities (performance, generality, user-interface, and communications) improvements in computer architecture will have to come from combining many designs. This paper considers various alternatives in the design spectrum for parallel computer architectures (tightly coupled systems, SIMD array processors, MIMD processing ensembles, clusters, data flow machines, etc.), their hardware and software requirements, and their potential for the near future.

© 1985 Optical Society of America