Evaluating and Mitigating Neutrons Effects on COTS EdgeAI Accelerators

Abstract

EdgeAI is an emerging AI accelerator technology which is capable of delivering improved AI performance at both a lower cost and a lower power level. With the aim of implementation in large quantities and in safety-critical environments, it is imperative to understand how Single Event Effects (SEE) affect the reliability of this new family of devices, and to propose efficient hardening solutions. Through neutron beam experiments and faultinjection analysis of a COTS EdgeAI device, we are able to identify the device's SEE failure-modes, separate the error rate contributions of the device's different resources, and characterise the device's SEE reliability. During this analysis we discovered that the vast majority of single bit flips have no appreciable effect on the output. After this analysis, we propose a hardening solution which implements TMR in the device without changing its physical architecture. We experimentally validate this solution and show that we are able to correct 96% of the misclassifications (critical errors) with nearly-zero overhead.

High Energy and Thermal Neutron Sensitivity of Google Tensor Processing Units

“…At ChipIR, the lowest SDC cross section is measured for ResNet, a similar neural network as the one tested for the NeuroShield in [1]. Considering a 13n/cm 2 /h flux for atmospheric neutrons at sea level, the ∼ 19 × 10 9 cm 2 cross section of the ResNet neural network translates in about 270 FIT, very similar to the 10 2 FIT measured for the NeuroShield.…”

“…Recently, some works have discussed the reliability of EdgeAI devices to neutrons and protons, focusing specifically the Arduino NeuroShield [1], [2]. To the best of our knowledge, this is the first work presenting experimental data on Coral TPU devices error rate.…”

“…Thanks to the intrinsic approximate nature of CNN computation, most of the errors do not turn into system failures, i.e., they do not affect the CNN accuracy. This has been proved for GPUs [11], FPGAs [13], and NeuroShield devices [1], [2]. Unfortunately, despite the intrinsic approximate nature, the misdetections and misclassifications rates in CNN executed in modern computing devices are still too high to be employed in safety-critical applications [11], [12].…”

“…As EdgeAI devices are likely to be used at scale and in distributed systems, it is fundamental to investigate their reliability, in particular their neutron-induced error rate. Preliminary studies showed that, despite being small, EdgeAI devices have a not negligible neutrons-or protons-induced error rate [1], [2].…”

Reliability of Google's Tensor Processing Units for Embedded Applications

“…As EdgeAI devices are likely to be used at scale and in distributed systems, it is fundamental to investigate their reliability, in particular their neutron-induced error rate. Preliminary studies showed that, despite being small, EdgeAI devices have a nonnegligible neutrons-or protons-induced error rate (BLOWER et al, 2021;BREWER et al, 2020).…”