In this paper we characterise two embedded GPU devices from the NVIDIA Xavier family System-on-Chip (SoC) using a proton beam. We compare the NVIDIA Xavier NX and Industrial devices, that respectively target commercial and automotive applications. We evaluate the Single-Event Effect (SEE) rate of both modules and their sub-components, both the CPU and GPU, using different power modes, and we try for the first time to identify their exact sources using the on-line testing facilities included in their ARM based system. Our conclusion is that the most sensitive part of the CPU complex of the SoC is the tag array of the various cache structures, while no errors were observed in the GPU, probably because of its fast execution compared to the CPU part of the application during the radiation campaign.