Evaluation of Fault Mitigation Techniques Based on Approximate Computing Under Radiation

A software technique based on approximate computing and redundancy is presented to mitigate radiation-induced soft errors in commercial off-the-shelf (COTS) microprocessors. Approximate computing relies on the capability of certain applications to accept imprecise results to improve efficiency by sacrificing its results in a controlled manner. Our approach avoids the time overhead derived from hardening while preserving the detection and correction rate. Experimental results show that we can detect and correct silent data corruption (SDC) events improving the cross section up to
$160\times $
and keeping accuracy under control without compromising performance. In addition, an accuracy-aware layer is included to improve error mitigation and to provide a tradeoff between the number of tolerable errors and the necessary accuracy.