Analyzing the Influence of Memory and Workload on the Reliability of GPUs Under Neutron Radiation

Evaluating the impact of utilizing different GPU resources is crucial for gaining insights into the reliability of GPUs when exposed to radiation. In this study, we employed various versions of a microbenchmark to investigate the effect of different memory types on the performance of a low-power GPU integrated into the Tegra X1 (TX1) system on a chip (SoC) of a Jetson Nano board. Additionally, we explored the tradeoff between enhanced computational performance and the occurrence of failures over time by optimizing the utilization of GPU resources. Our findings demonstrate that maximizing the utilization of the device’s cores enables the completion of a greater number of computations without errors. By fully harnessing the computational potential of the GPU cores, we effectively increase the work that we can complete between failures. Moreover, we observed that the use of the different memory types has a significant influence on the overall reliability of the GPU. The outcomes of this research contribute to a comprehensive understanding of the interplay between GPU resources, irradiation effects, and reliability. This knowledge is instrumental in guiding the development of robust GPUs for applications in radiation-prone environments.