GPUburn: A system to test and mitigate GPU hardware failures

Abstract

Due to many factors such as, high transistor density, high frequency, and low voltage, today's processors are more than ever subject to hardware failures. These errors have various impacts depending on the location of the error and the type of processor. Because of the hierarchical structure of the compute units and work scheduling, the hardware failure on GPUs affect only part of the application. In this paper we present a new methodology to characterize the hardware failures of Nvidia GPUs based on a software micro-benchmarking platform implemented in OpenCL. We also present which hardware part of TESLA architecture is more sensitive to intermittent errors, which usually appears when the processor is aging. We obtained these results by accelerating the aging process by running the processors at high temperature. We show that on GPUs, intermittent errors impact is limited to a localized architecture tile. Finally, we propose a methodology to detect, record location of defective units in order to avoid them to ensure the program correctness on such architectures, improving the GPU fault-tolerance capability and lifespan.