An Investigation into Fault Detection and Correction in GPU Pipelines with a Hybrid XOR Approach

Graphics Processing Units (GPUs) are continually finding new applications due to their broad parallel execution architectures. However, in safety-critical areas such as avionics, where hostile environments are common due to the effects of cosmic ray radiation, component failures can be highly problematic. This paper implemented, improved, and evaluated a hybrid fault tolerance technique initially proposed by NVIDIA to detect errors in the instruction execution flow against radiation effects affecting the pipeline of a GPU. Our results show that the technique can effectively detect all faults injected in the data-flow pipeline circuitry with an average 66% increase in execution time. Additionally, we were able to correct all detected faults with negligible costs.