Toward high-level synthesis of reliable circuits through low-cost modulo shadow datapaths

Abstract

With transistor dimensions shrinking to the atomic scale, a plethora of new reliability problems presents a barrier to continued Moore’s law scaling. Traditional modular redundancy techniques with 2x and 3x area cost eliminate the area reduction benefits of such scaling. In this study, we take a partial redundancy approach to the reliability problem for arithmetic-orientated datapaths by performing lightweight shadow computations in the mod-b space, where b is the base of our modulo residue, for each main computation. We leverage the binding and scheduling flexibility of high-level synthesis to detect control errors through diverse binding and minimize area cost through intelligent checkpoint scheduling and modulo-b reducer sharing. We introduce logic and dataflow optimizations to further reduce cost. We evaluated our technique with 12 high-level synthesis benchmarks from the arithmetic- oriented PolyBench benchmark suite using FPGA emulated netlist-level error injection. When b = 3, we observe coverages of 99.2% for stuck-at faults, 99.5% for soft errors, and 99.8% for timing errors with a 25.7% area cost and negligible performance impact. When b = 5, we observe coverages of 99.4% for stuck-at faults, 99.8% for soft errors, and 99.9% for timing errors with a 48.5% area cost and negligible performance impact. Leveraging a mean error detection latency of 13.92 and 14.96 cycles, with both mod-3 and mod-5 units respectively (2554x faster than end result check) for soft errors, we also explore a rollback recovery method with an additional area cost of 28.0% for both cases, observing 411x increase in reliability against soft errors.