Video Stream Processors: A Cost-Effective Computational Architecture for Image Processing

摘要

This report evaluates the capabilities of a new class of image-processing systems, known generically as video stream processors (VSPs). VSPs are an outgrowth of image display technology; image data from the display memory are streamed at video rates through a digital-processing unit and back to memory for subsequent display or further processing. This architecture serially simulates a parallel-array processor and is capable, in principle, of executing any locally parallel operation, such as convolution and edge detection, at a fraction of the cost of a truly parallel system. Our evaluation begins with a general discussion of the architecture and use of VSPs that highlights the fundamental concepts and vast application potential of this class of machines. This discussion is based on a hypothetical VSP design in order to avoid artificial constraints imposed by design limitations of any particular commercial product. The hypothetical design also serves as a standard against which current implementations can be evaluated. The report next summarizes our experience with the IP-5000 Image Array Processor, currently the most advanced commercially available VSP. The IP-5000 design is critiqued in the context of the hypothetical design, followed by a presentation of experimental results at SRI. The concluding discussion analyzes the IP-5000's limitations and proposes design refinements that would significantly improve its utility. (Author)