The STM32H7P is our first STM32 MCU to come with a hardware signal processor and a Neural-ART NPU, thus ensuring significant performance gains in intelligence and real-time control applications. It’s also our fastest STM32H7 with at least 1 MB of flash, since it runs at 520 MHz. It features up to 320 KB of Tightly-Coupled RAM, 192 KB of which is for instructions, ensuring deterministic execution of real-time code. Additionally, new timers open the door to digital power applications, while LQFP exposed pad packages help tailor systems for various industrial mission profiles. Despite all this, the STM32H7 has a similar unit price as the STM32H72x/3x series, thus seriously boosting the series’ price-per-performance ratio.
The challenge of bringing AI to existing MCU series #
What are most engineers looking for?
In a survey published by Aspencore in 2023, 655 respondents worldwide expressed that most embedded projects involve upgrading or improving an existing design and that real-time capability is the top priority. This explains why the STM32H7 remains a popular microcontroller series for engineers. It has broken records in the high-performance MCU market for more than a decade, often because ST optimized manufacturing processes to push its Cortex-M7 to run faster and with more cores, more memory, and hardware IPs. This has allowed teams to run more applications on one device while maintaining real-time performance, avoiding a more costly and complex microprocessor running embedded Linux.
Why are NPUs and HSPs great solutions with implementation challenges
However, new trends driving a rise in edge AI and physical AI applications are, in essence, a new challenge for silicon vendors because they push the boundaries of microcontrollers. A couple of years ago, we introduced the STM32N6, which comes with a dedicated NPU, the ST Neural-ART accelerator to significantly optimize AI algorithms. It has a lot of RAM but no flash, since engineers use external modules anyway. Similarly, to make our ultra-low-power series more capable, we launched a hardware signal processor on our latest STM32U3, which offers significant optimizations when it comes to fast Fourier transforms or other computations usually left to a DSP.
It is critical to support the next generation of applications that run neural networks or complex motor control. However, simply adding more IPs increases the die size, which makes a device vastly more expensive to manufacture. In some cases, it can also increase power consumption, which runs contrary to many projects’ requirements. Hence, ST had to rethink what the STM32H7 offered, which is why we came up with the STM32H7P, a device that comes with an NPU and an HSP without increasing the die size or its price, compared to the STM32H72x/3x series.
STM32H7P: The solution to bringing AI to a Cortex-M7 series #
Faster edge AI inference thanks to two AI hardware accelerators, without compromising performance
The NPU of the STM32H7P is a slimmed-down version of what we find in the STM32N6, with a throughput of 85 GOPS. Both devices serve different purposes. Pure AI applications will favor the STM32N6. However, adding a Neural-ART accelerator to the STM32H7P opens the door to new use cases, like an arc-fault circuit interrupter that can use a neural network to deliver performance unheard of on an STM32H7. In our tests, moving from the Cortex-M7 to the NPU meant that a solar panel could go from 212 AFCI model inferences to nearly 5,300. Concretely, that means managing more than a dozen additional solar panels with just one STM32H7P.
Similarly, the fact that we are using the same Hardware Signal Processor as the STM32U3 means that we can provide engineers a new path to optimizations akin to those available on newer architectures while keeping the same Cortex-M7. Indeed, in our labs, using the HSP of the STM32H7P to run matrix and vector operations often used in control loops, metering, or signal processing in digital power applications offers nearly identical results as if using the Cortex-M85 core with M-Profile Vector Extension of the STM32V8. Put simply, those who can’t move to our MCU with phase-change memory now have an avenue to benefit from similar performance when running certain computations.
Enabling deterministic real-time control and digital power even under load for stable and reliable applications
The STM32H7P includes 320 KB of tightly-coupled RAM and 320 KB of AXI RAM with ECC, which is important for AI algorithms. When the CPU, DMA controllers, or hardware accelerators need to move large amounts of data simultaneously, they provide higher bandwidth, burst support, and independent channels. We have already seen on other devices in this series, like the STM32H7R/S, how it is possible to address these memories without experiencing a wait state, thus taking advantage of the maximum clock speed. The new device also comes with up to 2 MB of dual-bank flash, so developers can switch between banks to update systems while keeping everything running.
The new device gets the most impressive set of timers ever on an STM32H7 series, with two high-resolution timers for up to 24 channels, three motor control timers, and four 12-bit ADCs for up to 44 channels. This enables developers to create more complex motor control or digital power applications. Those who need a more dedicated device will gravitate to the newly announced STM32G5. The strength of the STM32H7P is that it brings some of the same features that are so important for these new applications, but in a way that allows it to remain a general-purpose Cortex-M7 MCU.
Optimized to ensure thermal robustness in industrial designs without raising costs
The STM32H7P is the first in this series to offer exposed-pad packages, helping teams choose what works for their mission profiles. Thanks to the improved dissipation, the new device can support temperatures of up to 125 ºC at 350 MHz and 100% CPU utilization for 10 years. Many teams then refactor those numbers based on their mission profile to determine how the new device will behave, which will ultimately have a strong impact on the final bill of materials, since it means designers don’t have to plan for massive overheads. It also means that the STM32H7P can integrate a broader range of designs.
Another unique aspect of the STM32H7P is that despite all these innovations, ST managed to keep its die size identical to other devices in the series by optimizing the architecture for the customers that have already shown interest in it. Concretely, it means that there’s no Ethernet, display controller, or GPU. For instance, a partner has shown an interest in using the STM32H7P for audio processing in a car entertainment system. The new HSP enables entirely new applications that never need to run a GUI or that only need the device’s CAN-FD interface.
- Learn more about the STM32H7P