Critical Node Sensitivity Analysis and High-Reliability Design for CNN Accelerators on 16-nm FPGA Devices

Abstract

In recent years, Convolutional Neural Networks (CNNs) have emerged as a cornerstone technology for image processing tasks in safety-critical domains. Although advanced Artificial Intelligence (AI) acceleration devices are ideal for space applications due to their high computational efficiency and flexibility, they remain highly susceptible to Single Event Upsets (SEUs) and are still constrained by limited resources. Thus, it becomes imperative to design CNN implementations for Field-Programmable Gate Array (FPGA) devices that achieve a delicate balance between low cost and high reliability. However, traditional hardening strategies often neglect the sensitivity of internal nodes, relying instead on global or coarse-grained module-level hardening strategies that significantly increase resource overhead and degrade timing performance. To address this issue, we propose a basic unit partitioning strategy that divides CNN accelerators into functionally independent circuit nodes. Within our fault analysis framework, we map fault locations to specific nodes using address mapping, enabling sensitivity analysis. Additionally, we implement fine-grained hardening on critical nodes identified through sensitivity analysis and compare this strategy against module-level coarse-grained hardening and full triple modular redundancy (TMR) strategies. Fault analysis results reveal that our strategy delivers excellent hardening effectiveness at a reduced cost. Our strategy achieves a 31.9% reduction in resource overhead and a 29.9% improvement in timing performance compared to Full_TMR, with comparable fault tolerance. Compared to module-level hardening strategies, our approach achieves higher hardening efficiency. The irradiation experiment results further validate the effectiveness of our strategy, achieving a $10.6\times $ improvement in Mean Time To Failure (MTTF) compared to the unhardened baseline, which is comparable to that of full TMR. The proposed reliability evaluation method achieves high accuracy, with errors within ±4.3%. Overall, the proposed hardening strategy and reliability evaluation method present a significant advancement in designing robust and cost-effective systems, holding vast potential for adoption in safety-critical fields such as aerospace and autonomous driving.