Evaluating the Efficiency of ARM Cortex-M85 Architectural Extensions on Facial Recognition Machine Learning Workloads

This study explores the impact of different architectural options of the ARM Cortex-M85 microcontroller on the overall power efficiency in Machine Learning (ML) applications, specifically assessing its performance in facial recognition tasks across a number of hardware acceleration configurations. The increasing integration of Artificial Intelligence (AI) in embedded systems has created a demand for microcontrollers capable of executing ML workloads with minimal energy consumption. Our experimental approach is based on running a facial recognition model under various execution conditions, including cache and no-cache based memory hierarchies as well as DSP-assisted vector processing, while measuring power consumption for each inference. Our findings indicate that leveraging specialized acceleration units enhances power efficiency significantly, leading to lower energy consumption. Compared to conventional processing techniques of older microcontrollers, hardware-assisted ML execution modes demonstrated notable reductions in power usage while improving computational latency. These insights underscore the importance of advanced microcontroller architectures in enabling AI-driven applications at the edge, making them suitable for battery-powered and resource-constrained environments such as IoT and wearable devices.