Résumé
Convolutional Neural Networks (CNNs) enable autonomous on-board processing in space systems but remain computationally demanding and vulnerable to radiation-induced faults. This work evaluates the acceleration of quantized CNN inference on the fault-tolerant HARV-SoC, a RISC-V System-on-Chip, using its Vector Extension Unit (VEU) to vectorize convolution layers. Results from a proton irradiation campaign show that, compared to a related work on commercial microcontrollers, the HARV-SoC with the VEU achieves failure cross-sections that are up to two orders of magnitude lower.