Joint Optimization of Memory and Computing Frequency for Energy-Efficient DNN Inference

A new arXiv paper (2608.13863v1) proposes a joint optimization of memory frequency, computing frequency, and communication resources to minimize energy consumption in deep neural network (DNN) inference on mobile devices. The study addresses the overlooked impact of memory frequency on inference performance, formulating an optimization problem under deadline constraints. For local inference, a near-optimal closed-form solution is derived via convex optimization, and an optimal closed-form solution for transmission power is obtained. The research aims to enable energy-efficient DNN inference by considering both computing and memory frequency scaling, a departure from prior work that focused solely on dynamic voltage and frequency scaling (DVFS) for computing frequency.