Analysis and Mitigation of Soft-Errors on High Performance Embedded GPUs

Abstract

Multiprocessor system-on-chip such as embedded GPUs are becoming very popular in safety-critical applications, such as autonomous and semi-autonomous vehicles. However, these devices can suffer from the effects of soft-errors, such as those produced by radiation effects. These effects are able to generate unpredictable misbehaviors. Fault tolerance oriented to multi-threaded software introduces severe performance degradations due to the redundancy, voting and correction threads operations. In this paper, we propose a new fault injection environment for NVIDIA GPGPU devices and a fault tolerance approach based on error detection and correction threads executed during data transfer operations on embedded GPUs. The fault injection environment is capable of automatically injecting faults into the instructions at SASS level by instrumenting the CUDA binary executable file. The mitigation approach is based on concurrent error detection threads running simultaneously with the memory stream device to host data transfer operations. With several benchmark applications, we evaluate the impact of soft- errors classifying Silent Data Corruption, Detection, Unrecoverable Error and Hang. Finally, the proposed mitigation approach has been validated by soft-error fault injection campaigns on an NVIDIA Pascal Architecture GPU controlled by Quad-Core A57 ARM processor (JETSON TX2) demonstrating an advantage of more than 37% with respect to state of the art solution.