The most important part of the ML applied to MCU is to decide on a model. Model selection depends on what problem we are trying to solve. The easiest part is to use the existing model and it is already trained for a particular scenario. Even though the existing model is used, it is desirable to understand the architecture of the model and how it works on the data. At the end, the models have to run on devices. The device can be a full-fledged GPU (RTX 4090 for budget deep learning or the RTX 3060 for beginners) or the MCU we are talking about here. MCU have hardware accelerators which are specialized co-processors designed to perform specific, often computationally intensive, tasks more efficiently than the main CPU. These are built to handle the matrix and vector calculations common in AI and machine learning tasks like image recognition or voice assistants. Examples include the Arm Ethos-U NPUs found in devices from Alif Semiconductor or the STMicroelectronics Neural-ART Accelerator. Whether it be object detection, gesture recognition, AI/ML accelerators run inference on trained ML models. Ultimately, the target is to improve the performance of the MCU.
Examples in Modern MCUs:
NXP i.MX RT: Neural network and graphics accelerators
ESP32-S3: Vector AI/ML acceleration (for CNNs and FFTs)
Silabs SiWx917: Has MVP which acts a AI/ML accelerator
To summarize, modern microcontrollers are no longer just small CPUs, they come equipped with hardware accelerators that boost performance for specialized tasks. These accelerators are dedicated circuits designed to handle operations like signal processing, encryption, floating-point math, or even AI inference much faster than the main core could. By offloading these complex computations, accelerators help achieve real-time performance, reduce power consumption, and free the CPU for other control tasks. Common examples include FPU, DSP, Crypto engines, DMA, and AI/ML accelerators found in today’s MCUs like STM32H7, ESP32-S3, and Renesas RA series. They’re the hidden powerhouses enabling smarter, faster, and more efficient embedded systems.
0 Comments