Energy-Efficient Computing Acceleration of Unmanned Aerial Vehicles Based on a CPU/FPGA/NPU Heterogeneous System

The time and energy optimization of computationally intensive tasks involving unmanned air vehicles (UAVs) is highly important for increasing the reaction speed of UAVs and for prolonging their lifetime. To achieve the above objective, many studies based on heterogeneous computing have been carried out. Although these studies have achieved good results, limitations remain. First, neural processing units (NPUs) have emerged in recent years. However, insufficient attention has been devoted to CPU/NPU research in academia currently. Second, most popular heterogeneous computing architectures have only one kind of accelerator, e.g., CPU/GPU or CPU/field programmable gate array (FPGA). A heterogeneous system with multiple kinds of accelerators, e.g., CPU/FPGA/NPU, has not been investigated in depth. To address the above concerns, we propose a heterogeneous CPU/FPGA/NPU system aimed at realizing energy-efficient computing acceleration for computationally intensive UAV tasks. First, we select several representative computationally intensive UAV tasks and design FPGA and NPU accelerators dedicated to these tasks. Then, we calculate the time and energy costs of these tasks on the FPGA and NPU, respectively, and find that different tasks are appropriate for running on different cores. Based on this finding, we further build a heterogeneous CPU/FPGA/NPU architecture and assign each UAV task to the most appropriate core for execution. In this way, the UAV tasks can be executed more efficiently. We conduct experiments by executing all the representative UAV tasks on the CPU, CPU/GPU, CPU/FPGA, CPU/NPU and CPU/FPGA/NPU platforms. The results show that a heterogeneous system with multiple accelerators can achieve better computing performance and higher energy efficiency.