Yes, an SIMD Machine Can Be Used for AI

Abstract

Nearly all of the proposed parallel architectures for artificial intelligence applications use a multiple instruction, multiple data stream (MIMD) approach. While this approach offers the greatest opportunity for ultimate exploitation of parallelism, it has been difficult to actually achieve a high level of parallelism in practice because most of the existing algorithms require a higher interprocessor communications bandwidth than the hardware can achieve. An alternate approach to parallelism is the use of single instruction, multiple data stream parallelism (SIMD) machines. While SIMD machines do not offer the same ultimate exploitation of parallelism as MIMD architectures, they may, in fact, provide more useable parallelism. This is because they can be treated as serial machines with very long data words, so that existing algorithms may be more readily adapted in ways which better use available parallelism. We illustrate this concept with the Cellular Array Processor (CAP) being developed at the ITT Advanced Technology Center. This SIMD architecture is based on a rectangular processing array which accesses a single memory and which has very high speed data paths among the processing elements. We discuss the implementation and manipulation of data for two applications on the CAP; the OPS5 production system interpreter and a representation of an associative network. The expected performance of the CAP in these applications will also be presented.