The computing devices used in autonomous vehicles to detect objects and make decisions must be qualified to be highly reliable. Tests are necessary to measure the probability of failures caused by external perturbations, such as ionizing particles. These experiments are performed using particles accelerators such as the ChipIR facility of the ISIS neutron and muons source of the Rutherford Appleton Laboratory. The pandemic has prevent travels, but not the research!
Paolo Rech, that has recently joined the Industrial Engineering Department of the University of Trento, led a group of research that built a framework for the remote testing of neural networks accelerators. Using a private virtual network, researchers from Italy and Brazil were able to fully control the experiments from home. Besides looking at the data, the remotely controlled framework is able to detect when a reboot or a power cycle is needed (neutrons can stuck the device or cause the classical “Blue Screen of Death”) and perform the action. The results of experiments were very exciting, and allowed to understand how NVIDIA Graphics Processing Units and Google’s Tensor Processing Units behave when corrupted by neutrons. Two recent publications in IEEE Transactions on Nuclear Science and a presentation at the 2022 Design, Automation and Test in Europe Conference include some of the insights from the remote experiment.