Hardware accelerated embedded machine learning
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Hardware acceleration is a common method of improving the performance of algorithms running on general purpose processors. However it is not a widely used method for improving algorithms on embedded microcontroller units. This project investigates the performance gain associated with adding a hardware accelerator to a microcontroller unit (MCU) for a machine learning inference task using different scenarios. This work focuses on comparing the performance of four different scenarios: an 8-bit MCU only, an 8-bit MCU with a hardware accelerator, a 32-bit MCU only, and a 32-bit MCU with a hardware accelerator. Research was conducted by performing benchmark tests on each scenario using a machine learning inference task. The results showed that the addition of a hardware accelerator significantly improved the performance of the 8-bit MCU for the inference task and decreased performance for the 32- bit MCU. In conclusion, the study found that lower-end MCUs tend to gain more from a hardware accelerator than higher-end MCUs for machine learning inference tasks. The research has practical implications for developers and engineers working with MCUs in the context of machine learning inference, as it suggests that adding a hardware accelerator can be a cost-effective way to improve the performance of lower-end MCUs.