On the evaluation of soft-errors detection techniques for GPGPUs

Abstract

Recently, General Purpose Graphic Processing Units (GPGPUs) have begun to be preferred to CPUs for several computationally intensive applications, not necessarily related to computer graphics. However, due to their complexity GPGPUs also show a relatively high sensitivity to soft errors. Hence, there is some interest in devising and applying software techniques able to exploit their computational power by just acting on the executed code. In this paper we report some preliminary results obtained by applying two different software redundancy techniques aimed at soft-error detection; these techniques are completely algorithm independent, and have been applied on a sample application running on a Commercial-Off-The-Shelf GPGPU. The results have been gathered resorting to a neutron testing campaign. Some experimental results, explaining the capabilities of the methods, are presented and commented.