Neural Network Simulation on the MasPar MP-1 Massively Parallel Processor

Abstract

Progress in neural network R&D and technology transfer to real-world applications depends critically upon the availability of high-speed, massively parallel computing. The MasPar computer is a SIMD architecture well-suited for neural network implementations. We present an implementation of multi-layer perceptron learning using the back-propagation algorithm. The algorithm design is based on a virtual processor approach. An initial implementation of an unoptimized, general design yields approximately 306K connection updates per second (CUPS). This is 75% of previously reported results for an IBM 3090 (listed in Watanabe, et al., 1989, Table II) and 17-fold greater than our own Sun 3/80. We expect significant improvement to several 10’s of MCUPS as we begin to address load-balancing issues, better usage of the (1K) register space available to each processor and development of virtual processor depth reduction techniques.