STM32 is categorized into three tiers based on its AI support capabilities, as follows:
Tier One: High-Performance AI Specialized Series (with Hardware Accelerators)
These series are equipped with dedicated hardware accelerators (NPU) designed for AI computations, offering efficiency and speed far superior to standard cores, making them the preferred choice for handling complex AI models.
| Series Name | Core Features | AI Performance (Typical Values) | Applicable Scenarios |
|---|---|---|---|
| STM32N6 | The first Cortex-M core MCU designed specifically for AI, integrating NPU | Up to 1.5 TOPS (trillions of operations per second) | High-performance real-time visual AI, multi-sensor fusion, complex audio-video analysis |
| STM32MP2 (MPU) | Dual-core Cortex-A35 (application processor) + Cortex-M33 (real-time controller), integrating NPU | Up to 1.6 TOPS | Mid to high-end scenarios requiring complex operating systems (like Linux) and advanced AI applications, such as industrial HMI, high-end gateways, medical imaging |
Tier Two: High-Performance Mainstream Series (Relying on Powerful Cores and DSP)
Although these series do not have dedicated NPUs, they can efficiently run optimized AI models thanks to their high-performance Cortex-M7/M33/M4 cores, large memory, and advanced DSP instruction sets. This is currently the most mainstream battlefield for STM32 AI applications.
| Series Name | Core Features | Applicable Scenarios |
|---|---|---|
| STM32H7 | Ultra-high-performance dual-core Cortex-M7/M4, high frequency, large memory (up to 1MB Flash, 564KB RAM), with double-precision FPU | Complex computer vision (like facial recognition), real-time multi-channel audio processing, high-performance motor control + predictive maintenance |
| STM32F4 | Classic high-performance series, Cortex-M4 core with FPU and DSP instructions, mature ecosystem, high cost-performance ratio | Audio classification (keyword wake-up), vibration analysis (predictive maintenance), medium complexity image processing |
| STM32F7 | Performance close to H7, Cortex-M7 core | Similar to H7, an upgraded version of the F4 series |
| STM32U5 | Next-generation ultra-low-power high-performance series, Cortex-M33 core, providing strong performance under low power consumption, high security | AI applications requiring long battery life (like advanced wearables), smart healthcare, energy-efficient industrial sensors |
Features: H7 and F4 series are the most common models in STM32 Cube.AI examples and community projects, rich in resources, making them ideal choices for getting started and deploying mainstream AI applications.
Tier Three: Mainstream and Ultra-Low Power Series (Suitable for Lightweight AI)
These series are based on Cortex-M33/M4/M3 cores. Although their absolute computing power is not as high as Tier Two, they can competently handle lightweight AI tasks through extreme optimization with the STM32Cube.AI tool, especially excelling in applications that are highly sensitive to cost and power consumption.
| Series Name | Core Features | Applicable Scenarios |
|---|---|---|
| STM32L4+ / STM32L5 | Low-power high-performance, Cortex-M4/M33 core, still offers good performance in low-power mode | Battery-powered AI applications, such as offline voice command devices, simple gesture recognition, sensor node AI processing |
| STM32G0 / STM32C0 | High cost-performance mainstream series, Cortex-M0+ core | Extremely lightweight AI applications, such as simple anomaly sound detection (yes/no detection), binary classification (normal/abnormal) |
| STM32WB / STM32WL | Series with integrated wireless capabilities (Bluetooth/LoRa, etc.) | Wireless AI sensor nodes that perform preliminary intelligent analysis of sensor data locally before wirelessly transmitting results |
Features: Running AI on these chips requires significant model simplification and quantization (e.g., from 32-bit floating-point to 8-bit integer), but their value lies in achieving “intelligence at extreme cost”.
How to Choose? A Simple Decision Process
-
Define Your AI Task:
- Complex visual processing (like facial recognition) -> Prioritize Tier One (STM32N6) or high-end models from Tier Two (STM32H7).
- Audio classification, vibration analysis -> Tier Two (STM32F4/H7/U5) is the best choice, balancing performance, cost, and ecosystem.
- Simple keyword wake-up, binary sensor event detection -> Tier Three (STM32L4, G0) is completely sufficient and can effectively control costs.
Use ST Official Tools for Evaluation:
- Check if the model can run on your target STM32 model (check RAM/Flash usage).
- Estimate the inference time.
- STM32Cube.AI: This is the most critical tool. You can first import your trained model (like .h5 or .tflite files) into STM32Cube.AI, which will automatically analyze and inform you:
- STM32 Model Zoo: ST provides a library of pre-trained models, indicating which STM32 series each model is suitable for, serving as a great reference.
Summary Table
| Tier | Core Capabilities | Representative Series | Key Selection Factors |
|---|---|---|---|
| Tier One (Specialized) | High-performance NPU acceleration | STM32N6, STM32MP2 | Handles the most complex edge AI models, pursuing extreme performance |
| Tier Two (Mainstream) | High-performance cores + DSP | STM32H7, F4, U5, F7 | The backbone of most AI applications, perfect balance of performance, cost, and ecosystem |
| Tier Three (Lightweight) | Low-power/low-cost MCU | STM32L4, G0, WB | Extremely sensitive to cost and power, very lightweight AI tasks (binary classification, etc.) |
In summary, we should ask, “Which model is suitable for my application?” It is recommended to start with evaluation boards of STM32F4 or STM32H7, utilizing STM32Cube.AI for rapid prototyping, which is the best starting point for exploring the STM32 AI world.