A multiprocessor machine for large-scale neural network simulation

The architecture of a multiprocessor machine designed specifically for simulating large digital neural networks is described. The single-input multiple data (SIMD) machine comprises a host computer for high-level control and software development, a network controller which has data distribution and control responsibilities, and multiple processor arrays which carry out most of the computation for the network. Each processor array comprises a high-performance reduced instruction set computer (RISC) as a controller and 20 processing elements, each of which consist of a custom VLSI floating point processor and 1.5 Mbytes of private high-speed memory. The peak processing rate for a single processor array is 500 MFLOPS which can be sustained for relatively long vectors