Reliability of FINN-Generated CNN Accelerators for Image Classification on SRAM-Based FPGAs Under Heavy-Ion-Induced Faults

This study examines the performance of two convolutional neural networks (CNNs) designed for aerial image classification in the presence of radiation-induced bit-flips. We modify these CNNs by adjusting parameters such as quantization and parallelism to facilitate their implementation using the FINN inference engine, which is optimized for the AMD/Xilinx field programmable gate arrays (FPGAs). The aim is to evaluate the impact of different quantization levels, network topologies, and architectural parallelism on area, computational performance, and reliability in the presence of soft-errors. Emulated fault injection and heavy ion irradiation were performed. The results indicate that the same CNN topology can exhibit up to a $2.7\times $ difference in mean fluence to failure (M $\Phi $ TF) by altering quantization and architectural parallelism. The findings demonstrate that higher dependability can be obtained by carefully combining a suitable CNN topology with optimized quantization and architectural parallelism.