Learning in systolic neural network engines

Reports the analysis of a range of training algorithms implemented on a linear systolic ring. The main tool used in this project has been an architectural simulator of one such neural network engine, TNP-the Toroidal Neural Processor. This simulator enables machine code implementations of training algorithms to be developed. In addition, there is associated software which enables instruction counts for different hardware implementations to be evaluated. The TNP is a linear systolic neural network accelerator engine. The results provide quantitative data to aid in determining the design requirements of such engines. This can be accomplished in one of two ways: by assessing currently available processing elements/controllers or by determining, at least to a first order, the performance estimation of custom-linked processing elements.