Feature: Embedded design
to build ML-based systems on microcontrollers. Te result is a new class of intelligent embedded systems capable of sensing, analysing and reacting to the world in real time – without requiring a cloud connection.
Hardware foundations for embedded ML In this article we will focus on microcontrollers built around the Arm Cortex-M processor family. Tey have long been the workhorse of embedded systems, powering everything from industrial sensors to consumer electronics. Earlier Cortex-M processors were designed primarily for control applications. But recent generations have had architectural enhancements, making them highly suitable for DSP and ML workloads; see Figure 1.
Development tools for machine learning on Cortex-M microcontrollers
By Trevor Martin, Arm Technical Specialist, Hitex
T
he idea of running machine learning (ML) algorithms on resource-limited microcontrollers would have seemed unrealistic not so long ago. For decades, ML demanded powerful desktop processors or cloud infrastructure. Yet, the landscape is changing rapidly. As Reinhard Keil,
Senior Director of Embedded Technology at Arm recently observed: “By 2030 most embedded devices will embrace AI capabilities”. Te question is no longer if embedded systems contain ML algorithms, but how engineers can get there. However, the emerging ecosystem is now enabling developers
16 May 2026
www.electronicsworld.co.uk
Cortex-M4 and Cortex-M7 Te Cortex-M4 and Cortex-M7 cores now have improved mathematical capabilities, including hardware floating point units and enhanced integer arithmetic. Tese features significantly improve performance for signal processing tasks like filtering, feature extraction and spectral analysis. Many embedded ML pipelines begin with DSP processing of raw sensor data, so these improvements laid important groundwork.
Cortex-M55 and Cortex-M85 More recent processors – the Cortex-M55 and Cortex-M85 – push things further with their Helium vector extension, also known as M-Profile Vector Extension (MVE). Helium provides multiple 128-bit vector registers, capable of performing parallel operations on 8-, 16- and 32-bit data. Tese vector operations dramatically accelerate the matrix and vector calculations used in neural networks. For many ML workloads, Helium can deliver order of magnitude improvements in performance compared with scalar code.
Neural processing units: Arm Ethos Arm has also introduced a family of dedicated neural network accelerators, known as Ethos neural processing units (NPUs). An Ethos NPU essentially consists of large arrays of multiply- accumulate units designed specifically for neural network workloads; see Figure 2. By executing many operations simultaneously, these accelerators can deliver enormous performance whilst consuming low power. Depending on the device, performance reaches 480× acceleration compared with a soſtware-only implementation. While the Ethos NPUs can work with any Cortex-M processor,
they are commonly paired with Cortex-M55 or Cortex-M85 processors. Ethos has now been licensed for use in SoC and microcontrollers by over 20 Arm partners, including Alif, Infineon, Himax, Nuvoton, Renesas, Racyics and RiseLink/Beken. Te result is a new generation of low-cost microcontrollers with sufficient computing power to support sophisticated DSP and ML pipelines directly on the device. Hardware is only half of the story, however. Making effective use of these capabilities requires a robust development ecosystem, too.
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44
