MultiRuler: A Multi-Dimensional Resource Modeling Method for Embedded Intelligent Systems of Autonomous Driving

Youqing Xu, Boyu Li, Zongwei Zhu, Weihong Liu, Gangyong Jia, Guangjie Han, Xi Li

Abstract: With the development of vehicular technology, autonomous driving has experienced explosive growth, making it an important and popular research field. In autonomous driving scenarios, autonomous driving tasks are often neural network-based intelligent tasks with different levels of complexity. These intelligent tasks usually have different requirements for hardware resources such as computational performance, memory capacity and power consumption. Additionally, the evaluation becomes even more challenging due to the different system architectures and resource configurations of hardware devices from different manufacturers. It is crucial to accurately estimate the resource requirements of different intelligent tasks on different hardware platforms in advance.To address this issue, this paper proposes a multi-dimensional resource modeling method based on deep learning techniques and constructs an evaluation model to predict the consumption of three resources: computing performance, memory usage and power consumption.Experiments demonstrate that for the performance, memory and power metrics, the evaluation model built using this method for predicting seven mainstream neural networks achieves average error rates of 6.78%, 8.86% and 7.89% respectively on the NVIDIA Jetson AGX Xavier platform, and the average error rates for the same metrics are 8.05%, 9.25% and 3.67% respectively on the Cambricon MLU220 platform.