Thermal Characterization of AI Applications on AI Accelerators-equipped Microcontrollers

We investigate the thermal characteristics of tiny ML devices with AI accelerators, to understand the thermal impact on wearable devices. We identify the unique characteristics of temperature measurement, temporal and spatial temperature variation, define two novel temperature metrics, saturation temperature and rate, and report these metrics for each compute component. For the analysis, we conduct various benchmarks and evaluate the thermal profiles of different AI applications and tasks. The results and insights lay an empirical foundation for the development of heat-safe wearable hardware.