In this paper we investigate the reliability of Google's Coral Tensor Processing Units (TPUs) to both high energy atmospheric neutrons (at ChipIR) and thermal neutrons from a pulsed source (at EMMA) and from a reactor (at TENIS). We report data obtained with an overall fluence of 3.41 × 10 12 n/cm 2 for atmospheric neutrons (equivalent to more than 30 million years of natural irradiation) and of 7.55×10 12 n/cm 2 for thermal neutrons. We evaluate the behavior of TPUs executing elementary operations with increasing input sizes (standard convolutions or depthwise convolutions) as well as eight CNNs configurations (SSD MobileNet v2 and SSD MobileDet, trained with COCO dataset, and Inception v4 and ResNet-50, with ILSVRC2012 dataset). We found that, despite the high error rate, most neutrons-induced errors only slightly modify the convolution output and do not change the CNNs detection or classification. By reporting details about the error model we provide valuable information on how to design the CNNs to avoid neutron-induced events to lead to miss detections or classifications.

We investigate the sources of Detected Unrecoverable Errors (DUEs) in Graphics Processing Units (GPU) exposed to a neutron beam. Illegal memory accesses and interface errors are among the more likely sources of DUEs. ECC increases the launch failure events. Our test procedure has shown that ECC can reduce the DUEs caused by Illegal Address access up to 92% for Kepler and up to 98% for Volta. Additionally, we analyze if the compiler optimizations can impact the DUE sources distribution for the Matrix Multiplication. We found that the machine codes generated by the different optimization levels can change the DUE source by no more than 24% on average.

No abstract