| （英） |
Federated Learning, which enables multiple devices to cooperatively execute machine learning, is expected to provide benefits such as data privacy protection, distributed machine learning processing, and reduced communication costs. In particular, there is an increasing number of cases where edge devices are used in outdoor environments, with promising applications in environmental monitoring fields. When used outdoors, edge devices may be powered by natural energy sources such as solar power, necessitating efficient system design under limited energy resources. However, quantitative evaluation of actual power consumption in edge devices for federated learning has not been sufficiently conducted, and power consumption characteristics during GPU usage remain unclear. In this paper, we construct a federated learning system using Flower, an open-source framework for federated learning, and measure and evaluate GPU utilization, temperature changes, and power consumption when using NVIDIA Jetson Nano as an edge device. In our experiment, we build a federated learning system consisting of one server and two clients and execute federated learning using the CIFAR10 dataset. From the evaluation results, we investigate the relationship between power consumption and GPU utilization as GPU temperature increases over time. Furthermore, we evaluate the energy requirements for 24-hour continuous operation of federated learning. |