Arm Pushes AI, ML to the Edge with the Compact, High-Efficiency Helium-Powered Cortex-M52

Arm has announced a push to bring artificial intelligence (AI) and machine learning (ML) to the very edge of the Internet of Things (IoT) with a new Cortex design built to take advantage of its Helium acceleration extensions: the Cortex-M52.

"So today AI is everywhere, but to realize the potential of AI for IoT we need to bring machine learning optimized processing to even the smallest and lowest power endpoint devices," Paul Williamson, general manager of Arm's IoT business division, told us during a pre-launch briefing. "It's only then that we can truly scale IoT and drive the further innovation and deployment that we think is out there."

Developed in collaboration with Arm China, the Cortex-M52 is designed as a logical stepping stone from the Cortex-M3 and Cortex-M33 — offering the benefit of a familiar development environment with the bonus of a major performance boost for on-device machine learning workloads. "The M52 delivers a 5.6x performance uplift for machine learning, and 2.7x uplift in digital signal processing [over the M33], matched by improved scalar performance and advanced memory interfaces that ensure it can be designed into appropriate systems," Williamson claimed.

"Today, a software developer looking to solve an embedded computing challenge is looking for both DSP and ML performance in one. And that's what they need to create these compelling new solutions with the power of AI. Now to do this in the past, a developer might need a combination of a CPU, a DSP, and maybe a neural processor or NPU. Meaning that they would have to build the hardware and, once it was built, they may have to write, debug, and tune code across multiple chips or multiple processes within a single design that might need three separate toolchains, compilers, [and] debuggers. With Cortex-M52, and with Helium technology, we're delivering ML and DSP features with a single toolchain, and that gives them the capabilities they need in a unified consistent environment."

In raw performance terms, the Cortex-M52 sits below the more powerful yet also more power- and space-hungry Cortex-M55 and the high-end Cortex-M85. This is balanced out by a reduction in footprint and power requirements — and, Williamson says, opens up new potential for smart edge devices. "I was given an example recently where you can use something like the M52 very, very efficiently, and that is where you're doing very low frame rate or single-image sensing using ML techniques," Williamson told us.

"They were monitoring pest detection on crops — they would do manual inspections driving around, and a single bug in a single region was a real problem. They want to warn farmers and deal with it as soon as possible, [and] by having a little battery-powered sensor with image capture capability you can do very low frame rate — like maybe once a minute, once every 10 minutes — scanning and sampling of the appearance of the leaf to look for the presence of these bugs. And that allows them to have an active network of sensors that can cover an entire region of vineyards without the need to have people daily testing."

"We see the Cortex-M52 addressing a wide range of smaller low-power applications, including predictive maintenance, motor control, power management, and even voice and gesture-led machine interactions," Williamson concluded, "as well as even markets like the medical sector and remote wellness monitoring."

Those product won't be appearing on shelves immediately, though: the core IP is available to license from Arm now, but the company isn't expecting to see it realized in silicon until some time in 2024. Pricing for these parts has yet to be confirmed, though Williamson suggested it could be as low as "the sort of dollar level, one or two dollars" when implemented in "a very minimal system with minimal memory, minimal footprint."