SIFI: AMD southern islands GPU microarchitectural level fault injector

Alessandro Vallero, Dimitris Gizopoulos, Stefano Di Carlo

Abstract: General Purpose computing on Graphics Processing Unit offers a remarkable speedup for data parallel workloads, leveraging GPUs computational power. However, differently from graphic computing, it requires highly reliable operation in several application domains. In this paper we present SIFI a reliability evaluation framework for soft-errors on AMD GPUs built on top of Multi2Sim, a micro-architectural level simulator. SIFI is capable of computing different reliability metrics by means of two different techniques: fault injection and ACE analysis. Experiments performed on a set of 14 GPGPU applications targeting the AMD Southern Islands GPU architecture show the capability of the tool and the potential of its use to support decisions about the best architectural parameters for a given application.