An Adaptive Defect-Tolerant Multiprocessor Array Architecture

Abstract

Recent trends in transistor technology have dictated the constant reduction of device size. One negative effect stemming from the reduction in size and increased complexity is reduced reliability. This thesis is centered around the matter of fault recovery, in the subject of device fault-tolerance, and graceful system degradation in the presence of hard faults. Using a sparing strategy to re-use functional pipeline stages of faulty cores, we take advantage of the natural redundancy of multi-cores. This is done by the incorporation of a re-configurable network in which the cores of the system sit upon and has the ability to re-direct the data flow from the faulty pipeline stages of damaged cores to spare functional ones. The implementation requires the absence of global signals and thus pipeline stage operation needs to be decoupled. We also develop the bi-directional switch required for the network and implement a 4-core working example of our architecture as proof of concept and to evaluate the design. The 4-core design can guarantee correct functionality with 75% of system non-functional, in the best case scenario. The Defect-Tolerant pipeline has overhead of 1.92% in execution cycles and 14.4% in terms of operating frequency, for our custom made stress-marks. Such a system implemented with un-pipelined interconnect would lead to a pipeline with 50% lower frequency and x 2.1 longer overall execution time when the system has no faults. With our architecture and pipelined interconnect the frequency overhead is reduced by 34% and the overall execution time cost by 28% in the full 4-core system. The total execution time overhead, for our stress-marks, in the complete system ranges from x 1.5 to x 3.8 compared to the baseline, depending on the number of defects in the array. The area overhead is around 69% and power consumption, without incorporating any advanced power saving technique, is estimated between x 4 to x 5 times higher compared to the baseline.