Researchers deployed an unprotected open-source Tensil neural network accelerator on the Zynq UltraScale+ MPSoC platform to perform ResNet-20 inference tasks. During the experiment, the device was exposed to a cumulative dose of $4.29 \times 10^{10}$ p/cm$^2$ under proton irradiation ranging from 20 to 58 MeV. There were seven workload interruptions: two instances of notebook process restarts, four board resets or reboots, and one power cycle. Two of these failures caused output damage, returning incorrect CIFAR-10 classification results; in one instance, 39 consecutive inputs returned non-pooling categories. Although no abnormalities were detected in kernel logs, memory tests, or sampled power monitoring, the planned reconfiguration ultimately resolved the deadlock issue. All nine failures occurred in the beam area of 4 cm in nominal size (exposing the SoC, LPDDR4, and board-level circuits), and not in the central field of the 2-cm SoC…