STMicroelectronics’ first microcontroller featuring its homegrown dedicated AI accelerator is now in mass production. The STM32N6 features STMicro’s NeuralART 600-GOPS (INT8) accelerator to support AI in consumer and industrial products that need additional processing at the microcontroller power and cost point. This part is expected to become one of the fastest products in the STM32 line to reach $100 million in revenue, according to Remi El-Ouazzane, president of the microcontrollers, digital ICs and RF products group at STMicro.

This new part, the first in STMicro’s multi-generation roadmap of AI accelerator equipped microcontrollers, has been designed for the AI “tsunami” coming to tiny edge devices, El-Ouazzane said.

“I don’t want to sound overly dramatic or hyperbolic, but I really do believe this is the third seminal moment [in AI adoption] because you’ve seen what’s happening in data centers and on devices, but it’s the first time this class of products lends itself to the MCU space,” he said. “This is the beginning of a long journey, we are going to open the door to use cases we’ve never heard of, and I think we’re going to open the door to a new era when it comes to what STM32 means to the industry. So for us, it’s a pretty important time.”

STMicro’s vision is AI-capable IoT devices connected to the cloud, including all its existing markets in industrial, consumer and automotive microcontrollers. However, this vision will require more compute than microcontrollers currently offer.

“We expect AI compute performance at the tiny edge to grow by a factor of 50 to 100 times in the next five years,” El-Ouazzane said. “The continuum of tiny edge to cloud is the area where the compute growth will be the biggest over the coming years.”

AI at the tiny edge is being driven by smaller, more efficient AI model development, the desire to reduce cost and power by taking AI out of the cloud, and by real-time applications that need low latency, he added.

STMicro sees three main application areas for the N6. The first is sensor processing for anomaly detection in applications like predictive maintenance. The second is enabling use cases for microcontrollers that today need a microprocessor or an application processor.

“The industry always has been and always will be in search of lower power and lower cost implementations of a given application,” El-Ouazzane said. “With the STM32N6, it’s creating a huge inflection point when it comes to making a piece of equipment, because you’re reducing the bill of materials but it’s also actually more power optimized.”

A third use case, offering advanced AI services without using the cloud, he described as the most intriguing, since it includes applications that haven’t been identified yet.

“Most of the applications are surprising us every day by the way our customers are using this product,” El-Ouazzane said. “It’s a good feeling, because it represents the quintessence of the fact that we are opening a new era for microcontrollers by embedding real AI performance into them.”

Hardware acceleration

Microcontrollers with AI accelerators are already on the market. NXP’s MCX N series offers around 5 GOPS from its eIQ Neutron homegrown accelerator. Others are using Arm’s Ethos micro NPU, which can be implemented between around 5 to 500 GOPS, depending on clock speed. Infineon’s PSoC Edge products, for example, implement the Ethos U85 around the middle of this range. STMicro’s STM32N6 at 600 GOPS has the most powerful accelerator seen in this class of device so far.

“We took a very opinionated view a few years ago to develop our own accelerator capability,” El-Ouazzane said. “It is not solely about developing a systolic array or a dedicated AI accelerator. In this class of microcontroller, memory is the number one priority when it comes to being able to deliver the goods at the right amount of power. Here, the combination of what we can do in RAM and non-volatile memory is essentially dimensioning your ability to go and provide [the AI] you need. It’s not so much about the AI acceleration itself, it’s about how you leverage your memory fabric.”

The STM32N6 features an Arm Cortex-M55 CPU running at 800 MHz, the highest frequency ever for an STM32 part. It also has the largest RAM of any STM32 to date at 4 MB. STMicro has included a high speed memory interface, image signal processor, MIPI interface and built in graphics support. Two versions of the N6 will be made (one with the NeuralART AI accelerator, one without) in order to offer flexibility.

The NeuralART accelerator in the N6 offers 600 GOPS INT8 AI acceleration, 600× more than the previous most powerful STM32 part, the H7 (which does not have dedicated hardware acceleration). It can run at 3 TOPS/W. The N6’s Cortex-M55 has Helium, Arm’s DSP extensions, for AI and DSP functions.

The next generation of STMicro hardware will introduce digital in-memory compute, which will improve performance 4×, and a future third generation should offer a further 10×, leveraging STMicro’s non-volatile memory and phase change memory assets more aggressively, El-Ouazzane said.

“We were very committed to going down the path of our own AI hardware acceleration, simply because the Von Neumann approach is [dead on arrival] in the long run,” he said. “Because to go and seek the additional order of magnitude improvements when comes to the amount of TOPS per Watt, you have to move to in-memory computing, so the system becomes way more sophisticated [in terms of] what you do in your in your AI accelerator, your systolic array, and how you leverage memory, be it RAM or Flash, to get to the next level of performance. It’s because of that belief that we went down the path of owning our own technology.”

Software toolchain

STMicro’s secret weapon is its all-important AI software stack, ST Edge AI Suite, which had 51,000 new project starts from customers in the first 9 months of 2024, El-Ouazzane said. The toolchain includes NanoEdge AI Studio (a no-code tool for time series sensor data processing with STMicro’s models) and STM32Cube.AI (for those with more AI experience wishing to optimize models and performance).

The suite also links to Nvidia Tao tools for training or fine tuning models. This toolchain is the most mature among STMicro’s MCU competitors and will now serve both the company’s accelerated (N6) and non-accelerated MCUs.

Wireless module

Wireless is a key part of making STMicro’s edge AI vision a reality, El-Ouazzane said.

The company is collaborating with Qualcomm to integrate Qualcomm’s wireless SoCs into the STM32 ecosystem, as the wireless chip giant seeks to reuse technology developed for smartphones in the IoT. The first module from this partnership is a combination WiFi/Bluetooth/Thread module, the ST67W611M1, tailored for compatibility with any STM32 MCU and fully integrated into STMicro’s toolchain. This module will complement STMicro’s existing multi-protocol Bluetooth Low Energy, ZigBee, Thread and Sub-GHz portfolio.

“Wi-Fi is a hard market, whose standards are being defined and driven by both the mobile and the networking industry, and where in the next decade, velocity of feature upgrade is going to be quite steep,” El-Ouazzane said. “This is what we are counting on Qualcomm for.”

The companies intend to build future modules that will cover cellular connectivity too, he added.

The module will be available to select customers in Q1 2025, with broader market availability coming in Q2. The N6 is available in volume now.

From EETimes