Neural Computing on a One Dimensional SIMD Array

Parallel processors offer a very attractive mechanism for the implementation of large neural networks. Problems in the usage of parallel processing in neural computing involve the difficulty of handling the large amount of global communication between processing units and the storage of weights for each of the neural processor connections. This paper wi l l discuss how massive parallelism in the form of a one dimensional SIMD array can handle indefinitely large networks in near real time by efficiently organizing the memory storage of weights, and input and output signals. Very little overhead time is used in communication of signals between processing units, and there is no idle time for any of the units. An advantage of SIMD array systems is that the arithmetic processing is done bit serially, wi th the result that trade-offs can be easily be made between the processor speed and precision of the signals and weights.