Fault Mode Analysis of Neural Network-based Object Detection on GPUs with Neutron Irradiation Test

Neural network-based (NN-based) object detection algorithms are often implemented and executed on Graph Pro-cessing Units (GPUs). Understanding the reliability and fault modes of NN-based object detection is indispensable to improve and guarantee system reliability. In this work, we measure and analyze the fault modes of NN-based object detection running on GPUs using a quasi-monoenergetic neutron beam. Experimental results show that there are burst fault modes that repeat the same silent data corruption (SDC) errors and induce variant SDC errors. While the repetitive burst-mode errors of 56% probably originate from upsets in NN model parameters, the root cause of the remaining variant burst errors of 44% is unknown.