High Energy and Thermal Neutron Sensitivity of Google Tensor Processing Units

Abstract

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.

Neutrons Sensitivity of Deep Reinforcement Learning Policies on EdgeAI Accelerators

“…Also, we tested only object detection neural networks, while we explored the reliability of reinforcement learning models. In [24], the authors report the Edge TPU reliability under high energy and thermal neutron radiation. The authors tested several different types of classification and object detection neural networks, while in this work we use the same device and the same radiation type but report the reinforcement learning reliability of 4 different models.…”

Special Session: Reliability Assessment Recipes for DNN Accelerators

“…By performing radiation experiments, recent studies have demonstrated that DNNs accelerators are highly susceptible to transient faults induced by radiation [31], [45]- [47]. Recent research data suggests that while are mostly affected because of the high amount of available resources [48], [49] and the possibility of having multiple output elements corrupted, which undermines DNN reliability [46], [50], FPGAs, are mostly affected due to the programable hardware characteristics, where a transient fault can change the configuration memory and change the circuit [51]- [53].…”

Testability and Dependability of AI Hardware: Survey, Trends, Challenges, and Perspectives

“…Software-level [57], [58], [61]- [63], [66]- [91] RTL-level [92], [93] Microarchitectural-level [94] Gate-level [95]- [98] Transistor-level [99], [100] Chip-level [101]- [110] Radiation experiments [81]- [85], [111]- [114] Memristor crossbar-based architectures [115]- [118] oretical study is presented for Feed-Forward Neural Networks (FFNNs) deducing the number of failing neurons and synapses an FFNN can tolerate.…”

“…Finally, results on the reliability to neutrons of Google Coral TPU are reported in [114], considering elementary operations and several CNN models. It turns out that, despite the high error rate, most neutron induced errors only slightly modify the convolution output and do not change the detection or classification of CNNs.…”