Impact of Radiation-Induced Effects on Embedded GPUs Executing Large Machine Learning Models

Large machine learning (ML) models are able to outperform previous state-of-the-art ML models such as convolutional neural networks (CNNs). In this paper, we study the impact of radiation-induced effects on large ML models executing on the NVIDIA TX2 system-on-chip (SoC). In addition to characterizing the SoC hardware, we compare the radiation response of a popular CNN, ResNet-50, to two recent large ML models: Vision Transformer (ViT), and Data-efficient image Transformers (DeiT). Our evaluation includes experiments with both heavy ions and high energy protons at three different facilities. No SEL was observed in our heavy ion experiments with an LET of 37MeV/cm2 × mg at up to 80°C. Furthermore, we investigate how high energy protons and heavy ions at low LET affect the correct application output and limit the availability of ML applications. Our analysis reveals not only how radiation-induced errors propagate internally through the ML models all the way to the output, but also the main causes for fatal errors that lead to crashes in the operating system kernel. Particularly, the cross-sections measured for crashes were roughly 4× higher than the ones measured for silent data corruption (SDCs) for every application tested with heavy ions. Additionally, our evaluation of ML applications showed that general matrix multiplications (GeMM) presents a higher SDC cross-section than all the ML models tested (ResNet, ViT, and DeiT), roughly 21% higher with protons and 59% higher with heavy ions.