Pulsed Laser Single Event Effect on Deep Neural Network of Xilinx Ultrascale+ MPSoC

The article explores the susceptibility of a deep neural network implemented on a Xilinx Ultrascale+ multiprocessor system on chip (MPSoC) to single event effects (SEEs) under pulsed laser irradiations. The irradiation test involves scanning the MPSoC using a 1064 nm wavelength single-photon laser with three different energy levels: 5, 8, and 11 nJ. During these scans, three types of soft errors were detected: misidentification number varied, system halt (SH), and system reinitialization (SRE). The distribution of these soft errors across the MPSoC is determined. Notably, a significant burnout occurs when the current reaches approximately 1000 mA, and this phenomenon does not reoccur once the threshold current is set to 800 mA. Additionally, software fault injection is performed, and misidentification number variation and SH soft errors are also detected during this process. These findings confirm a possible source of the observed soft errors in pulsed laser irradiation. It suggests that single event upsets (SEUs) in the configuration memory (CRAM) could be a potential cause of the observed soft errors during laser scanning.