Design of low-power variation tolerant signal processing systems with adaptive finite word-length configuration

This paper concerns the design of low power digital signal processing integrated circuits in the presence of significant process variations. The basic idea is to leave smaller-than-worst-case timing margin for improving energy efficiency during the design phase and selectively reduce the finite word-length of circuit datapaths in post-silicon to eliminate all the timing faults during the run time. This simple idea can be intuitively justified by the fact that process variations may render only a few post-silicon datapaths to timing faults, while reducing the finite word-length of a few datapaths in signal processing systems may not necessarily make the overall algorithm-level performance unacceptable in run time. We present a design flow to implement this method and propose a dual finite word-length configuration strategy to simplify its real-life realization. Using linear low-pass filter and Turbo code decoder design at 45nm node as case studies, we quantitatively demonstrate that this adaptive finite word-length configuration design strategy may effectively relax the timing margin and accordingly reduce the power consumption by over 18% over conventional worst-case design approach.