An accurate analysis of the effects of soft errors in the instruction and data caches of a pipelined microprocessor

Instruction and data caches are well known architectural solutions that allow significant improvement in the performance of high-end processors. Due to their sensitivity to soft errors, they are often disabled in safety critical applications, thus sacrificing performance for improved dependability. In this paper, we report an accurate analysis of the effects of soft errors in the instruction and data caches of a soft core implementing the SPARC architecture. Thanks to an efficient simulation-based fault injection environment we developed, we are able to present in this paper an extensive analysis of the effects of soft errors on a processor running several applications under different memory configurations. The procedure we followed allows the precise computation of the processor failure rate when the cache is enabled even without resorting to expensive radiation experiments.