Error Mitigation Using Optimized Redundancy for Composite Algorithms in FPGAs

Error mitigation techniques have become mandatory in aerospace applications to cope with soft errors. Approximate error mitigation techniques are an attractive solution to reduce the overheads caused by conventional approaches, such as triple modular redundancy. These techniques use approximate redundant copies of the target design, which can be implemented with less resources, at the expense of slightly reducing the precision of the result in case of error. In this work, we propose optimized redundancy for composite algorithms, a novel hardening technique for algorithms that may be decomposed into more simple parts. Instead of adding identical redundant modules, we use complementary modules that can be composed to implement the target design. These complementary modules can also be compared to detect errors. However, when they are correct, they produce an exact result instead of an approximate result. The experimental results show that this technique can reduce the overhead and provide a better tradeoff between overhead and precision than existing approximate techniques.