01/01/1995 09:00 AM
Computer Engineering
Massively parallel processing is generally considered necessary to meet the long-range goals of artificial intelligence. Current commercial machines are optimized for numeric calculations, however, and custom machines are generally restricted to a single application or AI methodology. The MISC architecture extends SIMD architecture to better serve AI applications while retaining a general-purpose core. MISC is perhaps unique in its features resulting from an emphasis on search pruning and matching complex objects. These include multiple small arrays, high bandwidth access to large memory, an enhanced mesh communication network, and multiple special-function units. Analysis of the interaction between algorithm, architecture, and implementation technology demonstrates substantial speed improvement and cost-effectiveness. This paper discusses the MISC architecture and principles for the evolution of existing machines towards additional applications in knowledge processing.
