A Generalised Parallel Architecture for Image Based Algorithms

Abstract

Real time image generation and image understanding require levels of computing power, that are beyond that available from conventional sequential machines. Current commercially available systems aimed at this area make use of special purpose hardware to achieve the necessary throughput, but these systems can only achieve their performance for a restricted set of algorithms that are implemented in the hardware. A programmable general purpose parallel machine offers the possibility to achieve the required performance without restricting the choice of algorithm. Unfortunately it is by no means clear which parallel architecture should be used. Many general purpose parallel architectures have been proposed but none has proved universally applicable, their problem being that their performance tends to be highly dependent on the algorithms that are being used, and it is therefore difficult to claim any of them are truly general purpose. However parallel machines can still be highly effective in specific problem areas where the class of algorithm is known. Our aim has been to design a parallel machine that is optimised for image based algorithms in both graphics and image understanding. The architecture is not limited to a specific set of algorithms, but is instead optimised towards a class of algorithms which we believe are representative of image based algorithms. This has not been a paper study, but has resulted in us implementing such an architecture. We have achieved this by making use of industry standard components and integrating them into a system level architectural design. Also we have where possible used industry standard programming languages to program our machine.