Single event upset failure probability evaluation and periodic scrubbing techniques for hierarchical parallel vision processors

Abstract

This paper proposes novel single event upset (SEU) failure probability evaluation and periodic scrubbing techniques for hierarchical parallel vision processors. To automatically evaluate the SEU failure probability and identify all the critical elements in a processor, complementary fault injection methods based on logic circuit simulator and Perl script are proposed. These methods can be used to randomly inject faults into D flip-flops (DFFs) and various types of memory at the register transfer level (RTL) as well as to evaluate the vision processor performance. Based on the evaluation results, an accurate periodic scrubbing technique is proposed to increase the processor availability. The results denote that the peak availability of the processor over a period of one year can be improved from 18% to 99.9% after scrubbing the RISC program memory for a period of 10 4 s. Therefore, we can improve the fault-tolerance performance of a vision processor while avoiding unnecessary area and power costs using techniques ranging from evaluation to mitigation.