Feature: Embedded design
embedded systems typically implement a multi-stage processing pipeline. For example: 1. Sensor acquisition 2. Signal conditioning 3. Feature extraction 4. Machine learning inference. Two Arm libraries help manage this pipeline:
CMSIS-DSP CMSIS-DSP provides a modular collection of signal-processing functions optimised for Cortex-M processors. Its capabilities include filtering, FFT analysis, vector mathematics and matrix operations. Te library supports block-based processing, which improves efficiency when working with streaming sensor data. It also includes Python wrappers that allow DSP pipelines to be prototyped directly within Jupyter notebooks before deploying to embedded firmware.
CMSIS-Stream As pipelines become more complex, optimising data flow between processing stages becomes challenging. CMSIS-Stream addresses this by providing a Python configuration script and C++ state machine that optimise data flow through DSP and ML pipelines. Te framework helps developers to define processing graphs,
allocate buffers efficiently and ensure deterministic execution; see Figure 7. Tis is particularly useful in real-time embedded systems where memory and latency constraints are strict.
Testing and validation Once the system is implemented, careful testing is essential. Tis validation typically occurs at several levels during the model design process:
Model validation During the training process the available data is split into training, validation and test datasets. Te training process used the validation data to test the model aſter each training epoch, to give an indication of its performance. When the model is finalised, the unseen test dataset is used
to check the model’s final performance. Tis provides an early indication of its performance. To prepare the model for execution on an embedded device,
typically it will be quantised, moved from floating point to integer operation. Tis will result in a small loss of precision; we can use the same test data to measure this loss. Te final model can also be run in a Python version of the
on-device framework, to provide a close indication of its final performance.
DSP pipeline verification Signal processing stages should be verified independently, to ensure they produce correct feature vectors. Here the CMSIS-DSP library includes a Python wrapper that allows each of the DSP stages to be tested in a Jupyter notebook.
System testing Finally, the complete application should be tested on real or simulated hardware. Here the AVH can play a major role, first to confirm the correct performance of the algorithm and then enabling automated testing as part of a continuous integration system.
Looking over the horizon Embedded systems are entering a new era, where the pace of innovation in embedded ML shows no sign of slowing. Just as reaching the moon requires a powerful rocket, running
machine learning systems requires significant computational power. Now advances in processor architecture, neural accelerators and soſtware frameworks have made it possible to run sophisticated ML models directly on low power microcontrollers. Future Cortex-M processors will continue to increase in
computational capability whilst maintaining the low power consumption required for embedded devices. One particularly exciting development is the combination of
Cortex-M85 with the Ethos-U85 NPU. Tis configuration can deliver up to 4TOPS of ML performance. Perhaps even more significant is the growing support for Transformer architectures. Transformers – originally developed for natural language processing – have rapidly become the dominant architecture for many ML tasks, including vision and speech. Bringing transformer support to microcontrollers opens the door to an entirely new class of embedded ML applications. For embedded developers, the opportunity is enormous.
Te tools are ready, the hardware is available, and the soſtware ecosystem continues to mature. Te question is no longer whether embedded systems will become ‘intelligent’. Te question is: What will you build with them?
USEFUL REFERENCES: SDS
https://github.com/ARM-software/SDS- Framework
CMSIS-DSP
https://github.com/ARM-software/CMSIS- DSP
CMSIS-Stream
https://github.com/ARM-software/CMSIS- Stream
CMSIS-NN
https://github.com/ARM-software/CMSIS- NN
Ethos CMSIS Driver AVH
https://www.arm.com/products/silicon-ip- cpu/ethos/ethos-u55
https://arm-software.github.io/CMSIS_6/ main/Driver/group__vstream__interface__
gr.html
https://www.arm.com/products/ development-tools/simulation/virtual- hardware
www.electronicsworld.co.uk May 2026 21
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
