A generic sum of products parallel processor for neural networks and digital signal processing

In this dissertation a processor which targets algorithms using, as the core operation, the sum of products calculation is introduced and evaluated. The programmable parallel pipelined processor provides extremely good performance and flexibility and requires a small amount of hardware.;There are classes of algorithms and applications that when implemented on general purpose uniprocessors result in poor performance. One class in particular contains algorithms which use, as the core operation, the sum of products calculation. These algorithms are typically applied to large sets of data. Areas in this class include artificial neural networks (ANN) learning algorithms for ANNs, and digital signal processing (DSP).;A set of computational, communication, and storage requirements for general learning in ANNs and digital signal processing have been identified. A number of diverse algorithms for learning in ANNs and DSP are then programmed using the generic sum of products processor (GSPP) to show the flexibility of this processor. The processor performance for these algorithms has been evaluated using a novel evaluation approach as well as a simulator of the machine.;Applications such as NETtalk for ANN learning and 1-K discrete Fourier transform (DFT) for digital signal processing are used as benchmarks. These benchmarks have been used to compare the performance of the proposed GSPP to a number of other machines. GSPP performs extremely well when compared to other processors that require significantly more hardware.