Abstract
Neural networks (NNs) are now widely used for perception processing in autonomous systems. Data from sensors like cameras and lidars, after being processed by NNs, feed control algorithms that form the core of autonomy-related functions. Such NNs are implemented on graphics processing units (GPUs) and modern GPUs can be partitioned into multiple virtual machines, each implementing a separate NN. Given an autonomous system with multiple NNs, how should each NN be sized and the GPU implementing them be optimally partitioned? In this work, we study multiple GPU partitioning techniques with the goal of optimal and safe system-level control performance.