Evaluating the Effect of 14 MeV Neutrons on Google Coral EdgeAI

This work studies the influence of 14 MeV neutron radiation on a commercial of-the-shelf tensor processing unit. The device under test was a Google Coral USB accelerator for edge artificial intelligence operations running pre-trained models. Three different neural networks were investigated, two image classification models of different size and complexity and one semantic segmentation model. The errors in both software and hardware resulting from irradiation were assessed across all three models. It has been demonstrated that the cross sections of the soft errors depends on the model, where the larger and more complex models show higher cross sections. Nevertheless, all observed soft errors only led to minor changes in the output and no misclassifications or serious incorrect segmentations were observed. Furthermore, cross sections were calculated for hardware errors. Despite their smaller values than those associated with soft errors, they could result in complete hardware failure. These findings underscore the importance of a detailed comprehension of neutron radiation effects on edge devices to ensure the component's resistance to radiation-induced malfunctions. Subsequent efforts will prioritize a deeper understanding of the failure mechanisms in different models, aiming to enhance the durability and resistance of these devices when operating within challenging radiation environments.