[Editor's Note] On-device AI is expanding its application in industries, medicine, and smart homes due to its advantages of reducing cloud dependence and reducing delay, transmission costs, and power consumption. Driven by this expansion of applications, shipments of edge AI-equipped MCUs are expected to reach 1.8 billion units by 2030, and developers' choice of MCU is expected to become an important development point. Among them, STMicroelectronics' STM32N6 MCU is equipped with a proprietary Neural-ART Accelerator™ and enables real-time, high-resolution AI vision processing with a single MCU chip without a separate SoC. This can secure an advantage in the market by reducing component costs and simplifying design. Meanwhile, ST plans to present 'ST Edge AI Solution based on STM32N6' at the ' 2025 e4ds Tech Day' event to be held on September 9th. Accordingly, this magazine arranged an opportunity to hear about the 'STM32N6-based ST Edge AI solution' through an interview with ST Manager Moon Hyun-soo, who was in charge of presenting at this event.


■ Key features of ST Edge AI solution based on STM32N6

Edge AI refers to technologies that directly execute AI algorithms and machine learning models on devices such as embedded microcontrollers, microprocessors, and sensors for industrial and automotive applications.


Real-time data processing is possible because data collected from edge devices is immediately inferred using AI models, providing faster response times, improved data security, and higher bandwidth efficiency.

The STM32N6 is ST's latest high-performance microcontroller (MCU) optimized for edge AI implementation.

First of all, it is equipped with an Arm® Cortex®-M55 core operating at 800MHz and ST's proprietary NPU (Neural Processing Unit), Neural-ART Accelerator™.

Based on this NPU, it processes pre-trained artificial neural network models very efficiently.

Instruction fetch, decode, and executeTraditional embedded processor cores designed for sequential processing (execution) have limitations in executing AI models with optimized performance.

This is because neural network computational topologies often involve significant memory accesses and accumulation and multiplication operations, which are not optimized for traditional sequential architectures.

Therefore, a different architecture is needed to perform fast and efficient AI inference within the typical built-in constraints on power consumption and silicon area.

To address these requirements, neural processing units (NPUs) emerged.

NPUs are particularly well-suited for energy-constrained environments, such as microcontroller-based applications, due to their high efficiency.

It provides an optimal solution that can handle a variety of edge AI applications while maintaining low power consumption.

The integration of NPUs and microcontrollers significantly expands the capabilities of MCUs, enabling them to handle complex AI tasks that were previously impossible.

Traditionally, MCUs have been limited to simple AI applications such as low-resolution photo analysis, time-series analysis, or low frame rates due to their limited processing performance and energy efficiency.

But now, with the addition of NPUs, these microcontrollers can perform advanced AI functions such as fast-moving object segmentation, localization, pose estimation, object classification, and speech recognition.

Offloading AI inference tasks to the NPU frees the MCU to focus on other critical functions, enabling efficient real-time processing.

ST has been working on the development of Neural-ART accelerators since 2016, and the STM32Cube.AI software solution released in 2019 was also part of the ST Neural-ART Accelerator research and development at the time. Was influenced.

Since then, as we've seen the industry embrace STM32Cube.AI and engineers leverage our solutions to create innovative edge AI products, we've evolved the Neural-ART accelerator into a unique product.

Currently, no other general-purpose MCU manufacturer offers a hardware and software ecosystem for edge AI that is so customized and optimized.

The Arm® Cortex®-M55 core can accelerate vector and matrix operations required for signal processing (DSP) and AI model inference at the hardware level through Helium, an MVE (M-Profile Vector Extension) technology included in the Armv8.1-M architecture.

It also provides user-friendly online tools and software such as the 'ST Edge AI Suite' to help developers develop, evaluate, and deploy machine learning algorithms with the STM32N6 quickly and efficiently.

■ What do you think is the unique differentiating factor of this solution compared to competitors or alternative technologies?

High-performance edge AI implementation based on STM32N6 is possible through the following: △Execution of complex neural network models through Neural-ART NPU △Real-time processing based on hardware acceleration △Processing of high-resolution images and high-sampling frequency data △Parallel execution of various neural network models.

First, the STM32N6 integrates ST's proprietary Neural-ART Accelerator™, providing powerful on-device AI capabilities that can efficiently process a variety of artificial neural network models with a computing performance of up to 600 GOPS.

Especially with NPU operating at 1GHz Its major strength is that it can perform high-performance inference at very low power through the 800MHz Cortex-M55 CPU and support for vector operations based on Arm Helium.

Second, real-time image processing and inference can be performed on the STM32N6 MCU without an external separate image processing chip.

This is because it includes a built-in Image Signal Processor (ISP), an efficient pixel pipeline, and a MIPI CSI-2 interface.

This architecture, which integrates the NPU, ISP, CSI-2 interface, H.264 encoder, and NeoChrom graphics accelerator, is a highly differentiated solution in that it enables implementation of E2E (End-to-End) AI vision applications without an external SoC or dedicated processing unit, and can drastically reduce system complexity and bill of materials.

■ If you could give developers some hints on how to best use STM32Cube.AI or other SDKs for on-device AI,

To effectively implement on-device AI, it is recommended to actively utilize ST's STM32 Model Zoo GitHub, which is provided together with STM32Cube.AI.

The STM32 Model Zoo contains a variety of pre-trained artificial neural network models and example code optimized for the STM32, allowing you to quickly start evaluation and development without complex model configuration.

Additionally, by referring to the AI application examples (such as person detection and object classification) provided exclusively for the STM32N6, you can easily understand and apply the entire workflow from hardware setup to artificial neural network model conversion and real-time inference processing.

Using these materials, you can take on your first challenge with on-device AI. Developers can also build prototypes in a short period of time and naturally expand them into actual product development.

While we suggest using target device-specific tools, the ST Edge AI Core CLI version for all ST devices and the ST Edge AI Developer Cloud allow users to optimize and evaluate AI model performance on all ST hardware.

For STM32 MCUs, STM32Cube.AI (X-CUBE-AI) supports neural network optimization. NanoEdge AI Studio is an AutoML tool.

For STM32 MPUs, developers can use the STM32MP2 offline compiler for AI and Linux AI frameworks for OpenSTLinux (X-LINUX-AI).

For Stellar MCUs, StellarStudioAI is a software package for optimizing and deploying neural networks.

For MEMS sensors equipped with machine learning cores, data analysis, algorithm design, and model optimization can be performed using the online tool ST AIoT Craft and the desktop tool MEMS Studio. The MLC model zoo provides pre-optimized models.

For MEMS sensors equipped with ISPU, MEMS Studio supports data analysis and model optimization. ISPU model zoo provides pre-optimized models.

■ Why is the STM32N6 attracting attention in the on-device AI market, and how does it implement the key advantages of on-device AI over cloud-based AI?

As artificial intelligence (AI) becomes more pervasive and data collection grows exponentially, processing all data remotely in the cloud becomes unsustainable and impractical.

Broad adoption of AI across diverse products and applications requires more efficient local processing solutions—edge AI.

Solution developers can harness the power of AI right on edge devices with advanced microcontrollers featuring built-in neural network accelerators.

This approach offers significant benefits, including reduced power consumption, reduced network load, and reduced latency, enabling faster and more responsive AI-powered applications.

According to ABI Research, the edge AI market is expected to grow significantly over the next decade.

The data predicts a significant increase in the use of microcontrollers for edge AI applications across a range of industries, including agriculture, automotive, wireless networks, healthcare, manufacturing, personal and business devices, retail, and robotics, with the number of units expected to reach approximately 1.8 billion by 2030.

The STM32N6 is a product optimized for this market trend of on-device AI.

The STM32N6 dramatically improves DSP and AI computation performance by leveraging Arm Helium vector computing technology along with a high-performance CPU based on the Arm® Cortex®-M55 (800 MHz).

Here, ST developed its own designAn integrated Neural-ART Accelerator™ (1 GHz NPU, up to 600 GOPS) enables real-time inference with efficient power.

Additionally, the STM32N6 has strengths in high-resolution camera connection and image processing through the MIPI CSI-2 interface and built-in ISP, and is well-suited for vision AI and UI (User Interface)-centric applications as it integrates an H.264 encoder and NeoChrom graphics accelerator.

These features have made the STM32N6 a highly anticipated next-generation edge AI platform, enabling faster, safer, and more efficient AI capabilities to be implemented directly on-device, reducing dependence on the cloud, a traditional approach for industrial and consumer products.

Meanwhile, ST will participate in the ' 2025 e4ds Tech Day ' to be held at the ST Center on September 9th and present on the topic of 'ST Edge AI Solution based on STM32N6'. Applications for ' 2025 e4ds Tech Day' can be made on the official website ( https://www.e4ds.com/conference/techday/ ).

☞Continued from Part 2