A parallel processor for neural networks

Abstract

A deeply-pipelined digital parallel processor for the implementation of Multi-Layer Perceptrons is presented. It employs high-speed limited-precision integer arithmetic and allows good recognition performance in combination with a novel training algorithm. Internal dynamic RAM is provided for storage of the weights. The chip achieves a performance of 600 million multiply-and-accumulate operations per second and requires a silicon area of 70 mm/sup 2/ in a 1.2-/spl mu/m CMOS technology.