Performance Evaluation of Split Computing with TinyML on IoT Devices.

The proliferation of Internet of Things (IoT) devices has sparked a growing demand for lightweight and energy-efficient machine learning solutions, leading to the emergence of Tiny Machine Learning (TinyML). This paper presents a thorough evaluation of TinyML, encompassing its performance metrics, challenges, and prospects, focused on the use of Split Computing. Split Computing allows to offload a subset of layers of a neural network to a more powerful Edge server, to achieve a faster computation hence lower inference latency. We evaluate our proposal on a real testbed with ESP32 microcontrollers with different neural network structures, highlighting the benefits of split computing for IoT devices with varying conditions. Our results indicate that split computing on IoT devices is viable and can bring benefits particularly in heavy load scenarios where the network conditions may rapidly change.