AI is gaining incredible momentum, and it is making its way into the very heart of our embedded systems. This field, which we call edge AI, brings machine learning capabilities directly onto microcontrollers, microprocessors, and smart sensors.
While specialized chips like the STM32N6 series can run complex AI tasks, general-purpose STM32 microcontrollers are also capable of handling many applications, especially those involving time-series data from sensors.
Your STM32 is AI-ready
The data captured by sensors from vibration, current, or environmental monitoring represents an excellent opportunity to enhance your projects. By embedding a small AI model in your microcontroller, you can enable your device to perform intelligent monitoring, anomaly detection, or data interpolation locally. This adds significant value and can be implemented across the STM32 portfolio, from the ultra-low-power STM32U3 series and the mixed-signals STM32G4 series, to the high-performance STM32H5 series. Running AI algorithms enhances the core function of the MCU.
Consider the STM32U3 series as a prime example. This microcontroller is renowned for its exceptional ultra-low-power performance. We deployed various AI workloads on it and submitted the benchmark results to MLPerf Tiny benchmarking (round v1.3), and the outcomes speak for themselves.
Our NUCLEO-U385RG-Q board ran a keyword spotting workload at over 48 inferences per second while drawing
only 245mW. This remarkable efficiency is unlocked by the STM32U3's groundbreaking near-threshold design (a first for the STM32 family), which drastically reduces dynamic power consumption. This is a significant step forward as it cuts energy costs by nearly 6x compared with the STM32L4 and 2.5x compared with the STM32U5. For battery-powered IoT devices, this combination of high performance and ultra-low power finally eliminates the traditional trade-off between intelligence and battery life.
Beyond the STM32U3 example, we confirm our leadership in the field of edge AI with our broad portfolio of microcontrollers. From the smallest, ultra-efficient devices to the most powerful models, STM32 microcontrollers can integrate AI models directly on the device. For the most demanding applications, such as computer vision and advanced signal processing, we also provide MCUs with neural network accelerators, ensuring that developers can deliver intelligence at the edge without compromise.
A software ecosystem to make it happen
Capable hardware is only part of the equation: the software ecosystem is crucial for making AI integration seamless. Our latest STM32 edge AI software tools are a key enabler to simplify your design journey.
- NanoEdge AI Studio
If your project involves time-series data, this AutoML tool is a game changer. Simply import your sensor data, and within a few clicks, the studio automatically generates a bespoke AI library perfectly optimized for your target STM32. This tool provides the fastest route to transform raw sensor data into an on-device AI solution for tasks like anomaly detection or classification.
- STM32Cube.AI
For engineers looking to train and develop their own neural networks, you can turn to STM32Cube.AI. The tool seamlessly converts pretrained models from popular frameworks like TensorFlow Lite into optimized C code that can be deployed on any STM32 microcontroller.
- STM32 AI Model Zoo
The model zoo contains a variety of pretrained models for common applications that you can use as a starting point for your own projects. It offers services to finetune, quantize, and deploy your models using Python scripts, as well as application code examples to kick-start your project easily.
Edge AI can be deployed without specialized hardware. It is a high-potential technology that can enhance almost any STM32-based application. With our comprehensive hardware portfolio and user-friendly software tools, you can easily start building smarter, more efficient devices today.
Additional resource
- Solution overview - Edge AI solutions on STM32
First published on Oct 01, 2025
