Work In Progress: A New Task Model for Real-Time DNNs over GPU

Recently, deep neural networks (DNNs) have been utilized in real-time systems such as autonomous vehicles, where meeting temporal constraints is essential. However, executing such systems on CPU-GPU architectures can make scheduling analysis challenging due to the added delays caused by computing and memory contention. In addition, classic task models are not directly able to model accurately such systems. In this article, we propose a new task model called DNN Task Model (DTM). This model considers both DNN properties and GPU architecture at the same time. It allows us to distinguish between CPU and GPU tasks, provides information about the DNN application and give more accurate execution time analysis through consideration of data quality. We compute DTM from a source CUDA file and a set of real-time specifications of the system. The proposed model is extensible enough to be adopted to various DNN type applications allowing designer to compare candidate software and GPU architectures. Furthermore, we propose a graph optimization inspired by Tensor-RT.