STM32MP2: Embedded AI, Vision and Linux Processing for Industrial Systems

Embedded products increasingly need to combine Linux, real-time control, multimedia, connectivity, and AI in a single platform. Industrial cameras, machine-vision systems, human-machine interfaces, smart sensors, and edge devices often require more processing capability than a traditional microcontroller can provide while still needing low power consumption and deterministic control.

The STM32MP2 series from STMicroelectronics is a second-generation family of 64-bit microprocessors designed for industrial and connected embedded applications. The family combines Arm Cortex-A35 application processors with Cortex-M33 real-time processing, graphics, multimedia, security, connectivity, and—on selected devices—integrated AI acceleration.

What Is STM32MP2?

STM32MP2 is STMicroelectronics' second-generation STM32 microprocessor family. Unlike STM32 microcontrollers, which are primarily designed for bare-metal and real-time embedded applications, STM32MP2 devices provide application-class processing capable of running operating systems such as Linux.

The portfolio is divided into three major product lines:

STM32MP25x — higher-performance devices with extended connectivity and multimedia capabilities
STM32MP23x — cost-optimized devices derived from the STM32MP25 family
STM32MP21x — more power-optimized devices for essential MPU applications

This range allows developers to select an STM32MP2 device based on the processing, camera, graphics, AI, connectivity, and power requirements of the product.

Cortex-A35 and Cortex-M33 Processing

A major feature of STM32MP2 is its heterogeneous processing architecture.

STM32MP25x and STM32MP23x devices can include dual Arm Cortex-A35 processors running at up to 1.5 GHz, alongside a Cortex-M33 running at up to 400 MHz. STM32MP21x devices use a single Cortex-A35 with a Cortex-M33.

The Cortex-A35 processors are designed for application-level workloads such as Linux, networking, user interfaces, multimedia, and computer-vision applications.

The Cortex-M33 can handle real-time and security-sensitive tasks independently. This separation allows an embedded product to run a Linux application environment while maintaining a dedicated processor for deterministic control.

For example, an industrial vision system could use the Cortex-A35 for camera applications and AI processing while the Cortex-M33 manages system control, sensors, power states, or other real-time functions.

STM32MP2 for Embedded Vision

STM32MP2 also provides camera interfaces for embedded imaging applications.

Selected STM32MP2 devices include a MIPI CSI-2 camera interface with an integrated ISP camera pipeline. ST's graphics and multimedia documentation lists MIPI CSI-2 with Lite-ISP support for STM32MP21, STM32MP23, and STM32MP25 devices, with camera configurations supporting up to 5-megapixel capture at 30 fps on the listed interfaces.

A simplified camera pipeline can look like:

Camera Sensor → MIPI CSI-2 → ISP → Linux Camera Stack → Image Processing → AI / Computer Vision

This architecture allows a camera sensor to connect directly to the STM32MP2 platform rather than requiring a separate application processor for basic image processing.

For camera developers, this can simplify the overall system architecture while providing a path toward computer vision and edge AI.

Integrated ISP and Camera Processing

The ISP is particularly important when designing a camera system around STM32MP2.

Raw sensor data typically requires processing before it can be displayed or used effectively by computer-vision algorithms. The image pipeline can perform operations such as sensor data processing, color conversion, image enhancement, and other ISP functions.

ST documentation identifies the STM32MP2 camera architecture as including MIPI CSI-2 and ISP functionality, giving developers a hardware path from the camera sensor into the processing system.

This is useful for applications such as industrial cameras, smart displays with cameras, machine vision, inspection equipment, and intelligent edge devices.

Camera integration still requires careful consideration of the image sensor, MIPI CSI-2 configuration, clocks, GPIOs, regulators, Device Tree, Linux drivers, and ISP configuration.

AI Acceleration

One of the most important additions to STM32MP2 is hardware acceleration for AI workloads.

The STM32MP25x and STM32MP23x families include an integrated AI accelerator. On STM32MP25x, the NPU can deliver up to 1.35 TOPS when operated at 900 MHz. STM32MP23x provides an NPU with up to 0.6 TOPS at 400 MHz.

The AI hardware is integrated with the GPU architecture and uses the OpenVX software stack for neural-network execution.

This provides an alternative to running every AI operation on the Cortex-A35 CPU. A computer-vision application can instead divide workloads between the CPU, GPU, and NPU depending on the requirements of the model and application.

For example:

Camera → ISP → Preprocessing → NPU → Object Detection → CPU Application

This architecture can reduce CPU workload and make AI inference more practical for power-constrained embedded systems.

GPU and Multimedia Processing

STM32MP2 is not limited to AI and CPU processing.

Higher-end STM32MP25 devices include a 3D GPU supporting modern graphics APIs including OpenGL ES, Vulkan, OpenCL, and OpenVX. ST also provides video processing capabilities, including H.264 video encoding/decoding on supported devices.

This combination makes STM32MP2 suitable for products that need both computer vision and a graphical user interface.

An industrial machine, for example, could capture video from a camera, process the image, run an AI model, and display the resulting information on an operator interface using the same MPU.

Linux and the STM32 Software Ecosystem

STM32MP2 is designed to run Linux-based embedded systems through ST's OpenSTLinux ecosystem.

This gives developers access to a production-oriented Linux environment while retaining the STM32 ecosystem for low-level development and real-time control.

The STM32MP2 software environment also includes ST's ST Edge AI ecosystem. Developers can use ST Edge AI tools to select, optimize, benchmark, and deploy neural-network models to STM32 MPU platforms.

This creates a development path from AI model development through embedded deployment.

For camera products, Linux integration also means that developers need to consider the complete software stack, including sensor drivers, Device Tree configuration, V4L2, media-controller components, ISP configuration, GStreamer, and application-level computer vision.

Connectivity and Industrial Applications

Connectivity is another major focus of STM32MP2.

Depending on the device, the family supports interfaces including Ethernet, USB, PCIe, FDCAN, SPI, I²C, I³C, UART, SDIO, and display interfaces. Higher-end STM32MP2 devices can include multiple Ethernet ports and multiple CAN interfaces.

This makes the platform well suited to industrial equipment that needs to connect cameras, sensors, displays, controllers, and network infrastructure.

Potential applications include:

Industrial machine vision
Factory automation
Smart cameras
Industrial HMI systems
Edge AI devices
Robotics
Smart sensors
Connected industrial equipment
Vision-enabled control systems

STM32MP25 vs. STM32MP23 vs. STM32MP21

The different STM32MP2 families are intended for different system requirements.

STM32MP25 is the highest-performance option, combining dual Cortex-A35 processors, Cortex-M33, AI acceleration, 3D graphics, multimedia, camera connectivity, and extensive industrial connectivity.

STM32MP23 provides many of the same core architectural concepts in a more cost-optimized platform, making it attractive for embedded AI and connected industrial products where the highest level of connectivity is not required.

STM32MP21 focuses on essential MPU performance and lower power consumption. It can still provide MIPI CSI-2 camera connectivity and an ISP on supported devices, making it relevant to simpler embedded vision products.

Building a Camera System Around STM32MP2

Selecting an MPU is only one part of developing an embedded camera product.

The camera sensor, lens, MIPI interface, ISP, Linux driver, Device Tree, image-processing pipeline, AI model, and application software all need to work together.

A typical STM32MP2 vision system may therefore look like:

Image Sensor → MIPI CSI-2 → ISP → V4L2 / Linux → GStreamer or Application → NPU / GPU → Computer Vision

The exact architecture depends on the selected STM32MP2 device, camera sensor, image-processing requirements, and AI workload.

For products requiring custom camera hardware or specialized image quality, ISP tuning and sensor integration can also become major parts of the development process.

Why STM32MP2 Matters for Embedded Vision

The STM32MP2 family provides an interesting middle ground between traditional microcontrollers and higher-performance application processors.

It combines 64-bit Cortex-A35 application processing, real-time Cortex-M33 control, camera interfaces, ISP functionality, graphics, multimedia, security, connectivity, and AI acceleration in an industrial-focused platform.

For embedded vision products, this makes STM32MP2 particularly interesting when a system needs Linux and AI capabilities without requiring the processing complexity of a much larger edge-computing platform.

STM32MP2 provides a flexible platform for embedded vision products that need to combine camera processing, Linux, real-time control, connectivity, and AI within a single system. Explore our hardware design services to learn how we support camera integration, schematic and PCB design, and custom embedded vision hardware development around platforms such as STM32MP2.