Benchmarking of Microcontroller-based Platforms for EdgeAI Applications

The integration of EdgeAI into microcontroller-based platforms presents a significant advancement in the deployment of machine learning applications directly at the source of data collection. However, finding the right platform for a specific application, and its final implementation requirements, imposes the need of a prior characterization of a set of EdgeAI-enabled platforms. In this work, we present a benchmarking of three embedded platforms: Arduino Nano 33 BLE sense, ESP-EYE, and Sony Spresense; using four typical machine learning applications: image classification, anomaly detection, keyword spotting, and visual wake words. All implementations were developed using TensorFlow Lite Micro. For this characterization, we considered execution time (particularly, inference time) and power and energy consumption. The respective performance metrics of each task are also computed to guarantee there is no significant degradation in the machine learning model's performance. Our results showed that none of the models was significantly degraded in performance. Overall, the Sony Spresense platform had a lower average power consumption for all tasks, with a maximum average of 67.65mW for the visual wake words task, and presents the best execution time for three out of four applications, with up to 2.5× speedup with respect to the Arduino Nano 33 for the image classification application.