High Energy and Thermal Neutron Sensitivity of Google Tensor Processing Units

In this article, we investigate the reliability of Google’s coral tensor processing units (TPUs) to both high-energy atmospheric neutrons (at ChipIR) and thermal from a pulsed source [at equipment materials mechanics analyzer (EMMA)] reactor Thermal Epithermal Neutron Irradiation Station (TENIS)]. We report data obtained with an overall fluence $3.41\times {10^{12}} {\mathrm {n/cm}}^{2}$ for (equivalent more than 30 million years natural irradiation) notation="LaTeX">$7.55\times neutrons. evaluate behavior TPUs executing elementary operations increasing input sizes (standard convolutions or depthwise convolutions) as well eight convolutional neural networks (CNNs) configurations (single-shot multibox detection (SSD) MobileNet v2 SSD MobileDet, trained COCO dataset, Inception v4 ResNet-50, ILSVRC2012 dataset). found that, despite high error rate, most neutron-induced errors only slightly modify convolution output do not change classification CNNs. By reporting details about model, provide valuable information on how design CNNs avoid events lead misdetections classifications.