Testing permanent faults in pipeline registers of GPGPUs: A multi-kernel approach

Abstract

In the last decade, General Purpose Graphics Processing Units (GPGPUs) have been widely employed in high demanding data processing applications including multimedia and high-performance computing due to their parallel processing capabilities. Nowadays, these devices are considered as promising solutions also for high-performance safety-critical applications, such as autonomous and semi-autonomous vehicles. Current GPGPUs are designed targeting challenging execution requirements, e.g., related to performance and power constraints, forcing designers to use aggressive technology scaling solutions. Nevertheless, some implementation technologies are prone to introduce faults in the device during the operative life adding unaffordable effects and errors for the safety-critical domain. Hence, effective in-field test solutions are required to guarantee the target reliability levels. In this paper, we propose in-field test solutions based on Software-Based Self-Test (SBST) targeting the control-path of pipeline registers located in the Streaming Multiprocessor (SM) of a GPGPU. We resort to a multiple-kernel approach to detect permanent faults in these register fields. The solutions were designed employing NVIDIA CUDA, when possible, and lower level constructs elsewhere. Several usages and compilation restrictions are also described. Fault simulation results on an open-source VHDL GPGPU (FlexGrip) implementation of the G80 architecture of NVIDIA are reported, showing the effectiveness and limitations of the approach.