Arm Holdings PLC today unveiled a new microcontroller and an accompanying neural processing unit that, when used together, can allow “internet of things” devices to run artificial intelligence models up to 480 times faster than today’s silicon. 

The microcontroller, the Cortex-M55, joins the company’s ubiquitous Cortex-M family of processor designs. Arm said the chip will provide more AI performance than any other product in the series.

Cambridge, U.K.-based Arm doesn’t make any chips itself but rather sells blueprints under license to semiconductor manufacturers. The Cortex-M series is a family of blueprints for microcontrollers, tiny chips comparable to central processing units that are used in connected devices with limited performance needs. Processors based on Cortex-M designs have shipped with more than 50 billion systems ranging from wearables to medical equipment.

The M55 is based on the Armv8.1-M architecture the company unveiled early last year. The instruction set, the machine language in which the processor expresses computations, includes an extension called Helium that’s optimized for AI and digital signal processing. Helium has about 150 instructions specially geared toward these tasks along with a number of other performance optimizations.

The result, according to Arm, is that the M55 can run AI models up to 15 times faster than previous Cortex-M chips. The integrated digital signal processing capabilities, in turn, provide an up to fivefold performance uplift when it comes to ingesting data such as photos, videos and audio.

Hardware makers that need even more speed can deploy the chip in their devices together with the Ethos-U55, a neural processing unit Arm also debuted today. The product, which the company bills as a microNPU, is designed to accelerate the machine learning performance of the Cortex-M55 and other Cortex-M microcontrollers.

Arm said it boosts the M55’s AI processing speeds by a factor of up to 32. That enables it to deliver as much as 480 times better performance than previous-generation chips from the Cortex-M series.

The difference will likely be lower in real-world applications but the M55-U55 pairing still promises to provide a big boost for IoT devices.

A device running an AI voice assistant can, according to Arm, provide six times better “speed to inference” with the M55 than the Cortex-M’s family previous top end model, the M7. If the U55 microNPU is added in, the company says performance is up to 50 times better. And power efficiency improves by a factor of seven or 25 depending on whether the M55 is used alone or with the U55, respectively. 

“With these additions to our AI platform, no device is left behind as on-device ML on the tiniest devices will be the new normal, unleashing the potential of AI securely across a vast range of life-changing applications,” said Dipti Vachani, the head of Arm’s automobile and IoT business.

There are a few specific AI use cases the company has in mind. Currently, small IoT devices running Cortex-M chips are for the most part capable of only simple machine learning tasks such as detecting vibrations around their chassis or finding keywords in text. The M55 adds support for more sophisticated operations such as anomaly detection, while pairing the chip with the U55 will allow hardware makers to facilitate even more sophisticated use cases including gesture detection, speech recognition and certain biometric scanning features.

The first chips based on the designs are expected to hit the market in 2021. 

Photo: Arm