Using Hardware Performance Counters to support in-field GPU Testing

Nowadays, Graphics Processing Units (GPUs) have gained importance in several domains where a high computational effort is required (i.e., Artificial Intelligence). At the same time, their adoption extended even to domains (e.g., automotive, robotics, aerospace) where the effects of possible hardware faults can be extremely serious. Hence, it is crucial to identify methods allowing to quickly detect the occurrence of faults in a GPU while it works in the operational phase. In this paper we propose a method based on the adoption of hardware performance counters. We show that the method is able to detect permanent faults occurring in some critical modules, thus allowing to increase the overall reliability of the system. The method does not involve significant costs, since performance counters already exist in all real GPUs, e.g., to support silicon debug and can be easily accessed in software.