Highly parallel processors in military systems

Abstract

Parallelism has found its way into programmable processors as well as dedicated engines such as FFT and digital filters. However, choices of machine architecture are still open. We have evaluated two contrasting types to test their versatility and to compare their performance on algorithms related to military applications. Fine-grain SIMD, and coarse-grain MIMD machines (Mil-DAP and Transputer arrays) have been applied to a spectrum of problems including FFT, two-dimensional operators, associative processing, linear assignment, sorting, dynamic programming and ray tracing. These relate to military needs in spectrum analysis and image correlation, feature extraction from images, ESM, tracking with netted radars, speech recognition and terrain intervisibility. Possibilities for parallelism in combat simulators are also being examined. Each type of processor has been proved versatile and much more powerful than conventional sequential machines. DAP has the advantage on regular low precision algorithms, and on assignment and sorting operations where scatter, gather and shift operations are important. Transputer arrays will (when the T800 version is available) offer a better capability for floating point arithmetic and less regular tasks. The chief conclusion is that both architectures are extremely versatile and VLSIcompatible, and that choices between them will more often hinge on the cost of a minimal system, the quality of the development software and the portability of code, rather than on the fundamental properties of the machine topology.