Safety-Driven DNN Sizing for Vehicular CPS

Perception processing in cyber-physical systems (CPS) is now almost exclusively done using Deep Neural Networks (DNNs). Here, camera, radar and LiDAR data – in autonomous vehicles or robots – is fed into DNNs that detect surrounding obstacles and distances to them. These results are used by controllers to compute appropriate actuation signals. But a CPS typically has multiple state components, where each of them might be estimated using a different camera, radar or lidar and an associated DNN. Hence, an emerging problem is to implement multiple DNNs on a resource-constrained graphics processing unit (GPU). While many GPUs from NVIDIA and AMD allow them to be split into multiple virtual GPUs, there is little work on how to partition them, and therefore size the corresponding DNNs, when they are a part of the same CPS. In contrast to the existing practice of focusing on the inference accuracy of individual DNNs in isolation, we propose a system-level safety-driven DNN sizing (and hence GPU partitioning) scheme for vehicular CPS. Our main technical contribution is a detailed experimental evaluation of this DNN sizing approach and an empirical validation of the formal technique behind it.