Accurate Energy Measurements for Tiny Machine Learning Workloads

Benchmarking techniques play a crucial role to assess the performance of ultra-low-power tiny machine learning systems. The MLCommons framework provides procedures and tools to democratize the newfound technological capabilities and ensure wide adoption by sharing datasets and best practices. TinyML is a branch of MLCommons that specializes in integrating neural networks into embedded devices while priori-tizing factors such as inference time and power consumption, which are critical requirements given the limited hardware resources available. The inference time benchmarks allow us to understand the performance that a particular hardware can obtain. Conversely, the power consumption benchmark provides architectures, procedures, and tools for assessing the electrical current required by a device to perform inference operations. However, some issues related to the accuracy and precision of power consumption measurements seem to emerge. In particular, the architecture of the proposed trigger system exhibits some weaknesses concerning the time window for inferences. Moreover, the power consumption benchmarks are reported as single values, failing to provide uncertainty information. This study proposes a new architecture for the trigger system for the measurement method. Moreover, a prototype of a new experimental setup was realized and validated on STM32H7. Finally, power consumption results are provided with type A measurement uncertainty. Industries creating devices for embedded neural networks can apply the study's findings to make better design choices, especially by considering measurement uncertainty to determine the reliability of the results and the device under test.