Deep neural networks (DNNs) are widely used in various fields, but they suffer from high latency and energy overhead on resource-constrained devices. Most research has focused on dynamic voltage and frequency scaling (DVFS), a technique that balances latency and energy consumption by varying the processor's computing frequency. However, memory frequency scaling is often neglected or underutilized to improve DNN inference efficiency. In this paper, we use model- and data-driven methods to investigate the impact of simultaneously scaling memory and computing frequencies on inference time and energy consumption. Furthermore, we conduct a preliminary analysis of the proposed model by combining the fitting parameters of various DNN models and verify the effectiveness of simultaneously scaling memory and computing frequencies. Finally, we validate the effectiveness of jointly scaling memory and computing frequencies in reducing device energy consumption through local and collaborative inference simulation results.
