Optimizing Energy Consumption in Embedded and Optical Network Devices Using Trained Deep Neural Networks

Abstract

Power efficiency is now a decisive issue in developing embedded network hardware and optical networking devices, particularly when the concept of edge computing and always-on connectivity gains traction. This paper proposes our deep learning-based framework for dynamic energy optimization in distributed embedded systems and optical communication infrastructures. Based on optical training of deep neural networks (DNNs), our system can predict trends of device utilization and dynamically changing power states of components like optical transceivers, embedded processors, and programmable logic devices. The architecture leverages optical resources for the computationally expensive training exercise. It enables lightweight inference at the edge of the network devices, which is compatible with low-power network edge nodes. Validated on a system that uses microcontroller-based routers and optical switches, the system resulted in up to 18% reduction in average power consumed without hindering performance. Our findings present an opportunity for AI-driven energy policies to handle power budgets for innovative pro-environmental telecom and industrial purposes.