STM32H7P5: ST's First Cortex-M7 MCU With a Neural-ART NPU
ST's STM32H7P4 and STM32H7P5 pair a 520 MHz Cortex-M7 with a signal-processing engine, and the P5 adds an 85 GOPS Neural-ART NPU from $4.43 in volume.
STMicroelectronics has added two new chips to its high-end STM32H7 line, and they bring an unusual mix of features to a Cortex-M7. The STM32H7P4 and STM32H7P5 run at 520 MHz and include a Hardware Signal Processor (HSP) for maths-heavy work. The STM32H7P5 also adds ST's Neural-ART NPU, the first time ST has put that AI block in a Cortex-M7 part.
- Both chips run an Arm Cortex-M7 at up to 520 MHz, rated at 2,773 CoreMark
- The STM32H7P5 adds an 85 GOPS Neural-ART NPU for TinyML inference
- Memory is 1MB or 2MB of dual-bank flash with ECC, plus 736KB of SRAM
- Pricing starts at $4.43 for a 1MB part in LQFP100, in 10,000-unit quantities
What Does the Hardware Signal Processor Do?
The HSP takes FFT, matrix, vector, filtering and convolution work off the Cortex-M7 core. It supports 32-bit floating point, and ST says it runs FFTs up to five times faster than software on the CPU. The chips also keep the familiar CORDIC and FMAC maths accelerators.
That matters most in motor drives, digital power supplies and industrial sensing. In those systems the main loop has to do signal maths on a fixed schedule, and any time the CPU saves goes to control logic or communications.
How Fast Is the Neural-ART NPU on the STM32H7P5?
The NPU runs at up to 520 MHz and delivers 85 GOPS, or 0.085 TOPS. That is small next to the 600 GOPS NPU in the STM32N6 that powers the NeoEyes NE302 camera, but it is aimed at a different job. ST says its AI-based arc-fault circuit interrupter (AFCI) model runs up to 40 times faster on the NPU than in software on the Cortex-M7.
So a solar inverter or EV charger can run a fault-detection model and its real-time power control on one low-cost chip.
STM32H7P Timers, Peripherals and Packages
The peripheral list is built for control work:
- Two high-resolution timers with up to 24 PWM outputs and 1.9 ns resolution
- Three advanced motor-control timers and four 12-bit ADCs at up to 5 MSPS
- Three CAN FD controllers, USB OTG Full-Speed and two USB-C Power Delivery controllers
- Octo-SPI and FMC interfaces for external PSRAM, NOR, NAND or SDRAM
- Packages from an 8 x 8mm TFBGA100 up to LQFP176, rated -40 to +125°C
Security features include readout protection, active tamper detection, secure firmware upgrade and a hardware hash accelerator.
Development Boards and Availability
Software support comes through the STM32CubeH7P package with HAL2 drivers, STM32CubeMX2 and STM32CubeIDE. ST also offers motor-control and digital-power software packs. Two boards are ready for evaluation: the NUCLEO-H7P5II Nucleo-144 board and the STM32H7P5I-DK Discovery kit, which adds an OLED display, a MEMS microphone and a motor-control connector.
The chips are sampling to OEM customers now, and ST plans mass production for January 2027. If you have built on the ART-Pi II's STM32H7R7, the H7P is the AI-ready next step in the same family. Browse more boards and chips in our mini computer coverage.
Sources: STMicroelectronics newsroom — November 4, 2026; CNX Software — November 6, 2026; ST blog — November 3, 2026.
More Mini Computers Stories
M5Stack Core Matrix: A $29.90 ESP32-C61 Board With 256 LEDs
M5Stack's Core Matrix pairs an ESP32-C61 with Wi-Fi 6, a 16x16 white LED matrix and a BMI270 IMU for $29.90, ready for Arduino and UIFlow2 projects.
What Is Wi-Fi HaLow? 802.11ah Range, Speed and Hardware
Wi-Fi HaLow (802.11ah) reaches about 1 km on sub-1 GHz spectrum at up to 43.3 Mbps. Learn how it works, how it compares to LoRa, and which boards to buy.
LattePanda Mu Ultra Benchmarks: Local LLMs on a $733 Kit
CNX Software's LattePanda Mu Ultra benchmarks show 13.59 tokens/s on a 3B Llama model, 411 FPS YOLO26n on the GPU, and what the $733 kit needs for cooling.