Characterization of Sub-20 nm FinFET-Based AI Chips Under Single-Event Effects

Single-event experiments were conducted on four typical CNN architectures currently prevalent in the aerospace field: Yolov5s, Ssd, ResNet50, and Vgg16. During heavy ion testing, Single-Event Functional Interrupt (SEFI) phenomena were observed in all architectures. Among them, the Ssd architecture experienced more severe SEFI compared to the other three. Additionally, Single-Event Upset(SEU) phenomena were observed only in Yolov5s.To further investigate, pulsed laser experiments were carried out. The results showed that both Yolov5s and Ssd exhibited more severe SEFI, and SEU phenomena were detected in these two models. In contrast, the other two models only displayed SEFI without any SEU phenomena.By analyzing the differences in network layers and resource utilization among the four CNN architectures, potential reasons for these variations were summarized. This study provides a reference and insights for future in-depth research in this field, offering fault analysis and insights for the next generation of radiation-hardened AI chips.