Evaluating the effects of single event upsets in soft-core GPGPUs

Abstract

Graphic Processing Units have become popular in a broad range of applications due to their high computational power and low prices. Among the applications there are the safety critical ones, where fault tolerance is mandatory. This paper presents a fault injection methodology to evaluate a soft-core General Purpose GPU design over Single Event Upsets in its register files. Different configurations of CUDA algorithms are explored to verify their impact on the GPU's behavior during a fault injection campaign. This paper also presents an error characterization analysis by verifying the GPUs memories and program counters in order to evaluate the real impact of the fault in the GPU, even if the fault does not result in an error in the final output of the system. Results can help designers developing fault tolerant techniques in an effective fashion.