STMicroelectronics' STM32N6 Brings Its In-House Neural-ART NPU to Bear on TinyML, Computer Vision

Neural-ART, the company's first in-house coprocessor for tinyML, delivers 600 GOPS of compute alongside an 800MHz Arm Cortex-M55 core.

STMicroelectronics has officially launched the STM32N6, its first chip family to include a Neural-ART co-processor for tiny machine learning (tinyML) and on-device edge artificial intelligence (edge AI) workloads.

“We are on the verge of a significant transformation at the tiny edge," claims STMicro's Remi El-Ouazzane of the new chip family. "This transformation involves the increasing augmentation or replacement of our customers’ workloads by AI models. Currently, these models are used for tasks such as segmentation, classification, and recognition. In the future, they will be applied to new applications yet to be developed,.

"The STM32N6 is the first STM32 product to feature our Neural-ART Accelerator NPU. It will utilize a new release of our unique AI software ecosystem package. This marks the beginning of a long journey of AI hardware-accelerated STM32, which will enable innovations in applications and products in ways not possible with any other embedded processing solution.”

STMicro first unveiled the STM32N6 in a live demo at Embedded World last year, showcasing its capabilities in a head-to-head challenge against the STM32H747. Running a customized You Only Look Once (YOLO)-derived neural network, trained to locate people in live video, the STM32N6 delivered a 75× increase in performance — yet ran at less than half the clock frequency. After the show, the company offered a second comparison: the claim that the STM32N6 delivers 25× faster inference for on-device machine learning than the STM32MP1, a dual-core Arm Cortex-A7 application-class processor running at 800MHz.

"The STM32N6 redirects its AI compute task to the ST Neural-ART Accelerator and its preview functions to the STM32N6's Machine Vision pipeline," STMicro's Miguel Castro explained of the demo's smooth performance at the time, "leaving the Cortex-M with the flexibility to handle other tasks."

STMicro has demonstrated a range of models running on-device, accelerated by the Neural-ART coprocessor. (📹: STMicroelectronics)

The STM32N6 chips are based on an Arm Cortex-M55 core running at up to 800MHz, while the Neural-ART coprocessor runs at up to 1GHz and delivers a claimed 600 giga-operations per second (GOPS) of compute at a three tera-operations per second per watt (TOPS/W) power draw. Other coprocessors include a Chrom-ART accelerator for 2D graphics, a Chrom-GRC "graphics resource cutter" for round and other non-square displays, a "2.5D" NeoChrom graphics accelerator, an H.264 video encoder capable of 1080p15 or 720p30, and an image signal processor (ISP) targeting a five megapixel camera at 30 frames per second.

Interestingly, the chips do not include any on-board flash; instead, they offer a "flashless" memory configuration with 4.2MB of contiguous embedded RAM, plus external interfaces for multiple memory types including pseudo-static RAM (PSRAM), synchronous dynamic RAM (SDRAM), and both NOR and NAND flash. There's the Arm TrustZone security subsystem, a choice of side-channel-attack resistant and high-speed AES acceleration, tenant-aware firewalling, and the target of achieving SESIP Level 3 and PSA Level 3 security certifications.

STMicro has confirmed two families of STM32N6 will be available at launch: the STM32N6x7 range includes the Neural-ART coprocessor, while the STM32N6x5 range drops it for use as a general-purpose high-performance microcontroller; both ranges will also be available with or without the hardware cryptographic acceleration blocks. All models include Arm's Helium vector extensions, boosting machine learning workloads while running on the microcontroller core.

The new parts are now available "in high volumes," STMicro has confirmed, after sampling to select customers in October. A developer kit built around the STM32N6 has been announced at $185, with a NUCLEO development board at $56.25; more information is available on the STMicro website.

Gareth Halfacree
Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.
Latest articles
Sponsored articles
Related articles
Latest articles
Read more
Related articles