Bringing Edge AI to low-power nRF and STM microcontrollers
Jussinmäki, Lauri (2025-07-01)
Bringing Edge AI to low-power nRF and STM microcontrollers
Jussinmäki, Lauri
(01.07.2025)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe2025073080089
https://urn.fi/URN:NBN:fi-fe2025073080089
Tiivistelmä
This thesis investigates the development and deployment of embedded artifcial intelligence applications on low-power microcontrollers using the Edge Impulse platform. The work focuses on implementing a real-time keyword spotting system on two platforms: the Nordic Semiconductor nRF5340, and the STMicroelectronics STM32L476. These devices represent commonly used microcontrollers in batterypowered embedded systems without dedicated AI acceleration hardware. The study demonstrates a complete workfow from dataset creation to model training, optimization, and integration with real-time frmware using both Zephyr and FreeRTOS. The Edge Impulse platform was evaluated for its usability, performance estimation tools, and deployment options. Performance was measured in terms of inference latency, memory usage, and real-time feasibility.
Results show that meaningful machine learning tasks can be executed on generalpurpose low-power MCUs, with the nRF5340 ofering better performance due to its more modern architecture and higher clock speed. The STM32L476, while functional, showed limited headroom for real-time inference with the tested model. Edge Impulse’s performance estimations were found to be slightly optimistic for STM32 and slightly conservative for nRF5340, highlighting the importance of early on-device testing.
The fndings provide practical guidance for selecting microcontrollers for Edge AI applications and underscore the importance of early-stage performance testing. The thesis also outlines future directions for exploring newer MCU families and comparing general-purpose microcontrollers with systems that include dedicated AI accelerators.
Results show that meaningful machine learning tasks can be executed on generalpurpose low-power MCUs, with the nRF5340 ofering better performance due to its more modern architecture and higher clock speed. The STM32L476, while functional, showed limited headroom for real-time inference with the tested model. Edge Impulse’s performance estimations were found to be slightly optimistic for STM32 and slightly conservative for nRF5340, highlighting the importance of early on-device testing.
The fndings provide practical guidance for selecting microcontrollers for Edge AI applications and underscore the importance of early-stage performance testing. The thesis also outlines future directions for exploring newer MCU families and comparing general-purpose microcontrollers with systems that include dedicated AI accelerators.