Despite extensive research into improving fabrication processes, continued VLSI scaling will be inhibited by high variations in process parameters, higher defect densities, and higher susceptibility to external noise. We propose two notions, namely error-tolerance and acceptable operation, to facilitate imprecise computation: these notions systematically capture the fact that an increasingly large class of digital systems can be useful even if they do not perfectly conform to a rigid design specification. We propose to develop a systematic methodology for design and test of this class of digital systems that will exploit the notion of error tolerance, to enable dramatic improvements in scale, speed, and cost. In the proposed methodology, system specification will include a description of the types of errors at system outputs, and the thresholds on their severities, that are tolerable. The design methodology will exploit this information to obtain designs that provide higher performance and/or lower costs.
Over the next 15 years, the proposed approach will provide dramatic improvements in scale, speed, and cost for a wide class of digital systems, including many integral to NHS. This will enable development and wider deployment of devices with advanced capabilities in areas such as speech processing, real-time translation of spoken natural languages, and biometrics.
