Error Detection and Mitigation of Data-Intensive Microprocessor Applications Using SIMD and Trace Monitoring

Abstract

This article proposes a software error mitigation approach that uses the single instruction multiple data (SIMD) coprocessor to accelerate computation over redundant data. In addition, an external IP connected to the microprocessor's trace interface is used to detect errors that are difficult to cover with software-implemented techniques. The proposed approach has been implemented in an ARM microprocessor, and an irradiation campaign with neutrons has been carried out at Los Alamos National Laboratory. Experimental results demonstrate the high error coverage (more than 99.9%) of the proposed approach. The neutron cross section of errors that were not corrected nor detected was reduced by more than three orders of magnitude.