Abstract: With the rapid development of generative artificial intelligence technology, edge machine learning has gradually become a key trend, aiming to use data on edge devices for model training to reduce latency and improve user experience. However, due to the energy and resource limitations of edge devices, especially in executing high-energy learning tasks, improving the energy efficiency of edge devices has become the main challenge at present. This paper proposes an edge energy efficiency optimization model based on heterogeneous computing. On the local computing side, central processing unit and neural network processing unit (CPU-NPU) heterogeneous computing units are configured to reasonably allocate tasks between CPU and NPU to improve carrying capacity and energy utilization. On the transmission side, non orthogonal multiple access transmission is used to offload tasks from edge devices to servers to improve spectrum and energy efficiency. At the same time, in order to maximize the energy efficiency of edge devices, a joint computing and communication resource management optimization strategy is proposed. Based on the joint optimization algorithm, the single objective energy efficiency optimal solution for local heterogeneous computing task allocation, transmission power allocation factor, task offloading delay involving computing and communication sides is jointly optimized to maximize the overall energy efficiency of edge devices. The simulation results show that compared with traditional CPU single computing nodes, the proposed scheme improves spectrum resource utilization while also improving the energy efficiency of Internet of things (IOT) edge user devices by 30%.

Key words: heterogeneous computing, task unloading, edge computing, energy efficiency