With its new Cortex-M52 chip, Arm enables AI analysis on small IoT devices at the edge.
AI is everywhere these days, but where it’s really needed is in the farthest reaches of the edge, out where Internet of Things (IoT) devices are generating massive amounts of data that need to be analyzed, where predictive insights can be collected and acted upon, and where machine learning-optimized computing can be run.
“It’s only then that we can truly scale IoT and drive the further innovation and deployment that we think is out there,” said Paul Williamson, Senior Vice President and General Manager for Arm’s IoT Line of Business. “But developers face challenges. They need to extend their platform’s hardware capabilities but also a simplified software development platform.”
That’s what Arm says it is delivering with its new Cortex-M52, the latest addition to the chip designer’s family of low-cost and energy-efficient Cortex-M microcontrollers that is aimed at the smallest IoT and embedded devices, Williamson told reporters during a press briefing.
“The intersection of AI advancements and the pervasiveness of connectivity technologies means that on-device intelligence can be enabled in small, cost-sensitive devices and they can become smarter and more capable,” he said. “These devices can operate with great privacy and reliability due to less reliance on the cloud.”
Williamson added, “for the industry to realize that opportunity, silicon providers and developers need access to more AI capability to deliver the required intelligence, but do so in an area that’s cost- and power-constrained that’s typical of a small, embedded device.”
A Fast-Growing Market for AI in IoT
The AI-in-IoT sector is a rapidly growing space as organizations turn to artificial intelligence and machine learning to make sense of the data these devices — there are about 15 billion IoT devices now — are generating. Pragma Market Research expects that segment of the market, which generated $10.3 billion last year, to grow to $91.7 billion by 2032.
“This is core to the success of IoT, trying to get as much intelligence into the endpoints to reduce the loading on the related cloud services,” said Rob Enderle, Principal Analyst with The Enderle Group. “But it has to be very efficient as there often is very little power available at the endpoints, [including] battery and sometimes solar.”
Arm built the Cortex-M52 to include the company’s Helium technology, an extension of the Armv8.1-M architecture that improves performance for machine learning and digital signal processing (DSP) applications. With Helium, the Cortex-M52 will give a boost to both DSP and machine learning performance without dedicated DSP or machine learning accelerators or an Arm Ethos neural processing unit (NPU), which are seen in Arm’s high-end Cortex-M85 and midrange M55.
The Helium Advantage
Williamson said that the performance lift from Helium lets developers deploy more compute-intensive machine learning algorithms. The Cortex-M52 was created through a collaboration of Arm and an engineering team with Arm China and offers a home for workloads that are now running on the Cortex-M4 and M33.
It delivers a 5.6-times improvement in performance for machine learning jobs, a 2.7-fold performance jump for DSP tasks, and 2.1 energy efficiency compared to previous Cortex-M generations.
There also is a 23% reduction in the silicon area, giving silicon makers more choices when looking for trade-offs between performance and cost.
For security, the Cortex-M52 uses the latest extensions for the Armv8.1-M Pointer Authentication and Branch Target Identification (PACBTI) extension and Arm’s TrustZone technology. It also will help chip makers get to PSA Certified Level 2 silicon to create PSA-certified IoT devices.
Developers Get a Unified Environment
For developers, it opens up a wide range of use cases, including vibration, anomaly, and keyword detection as well as sensor fusion.
“It gives them more options when optimizing an IoT solution,” Enderle said.
The new chip design provides a unified software development environment and Cortex-M toolchain. Other Cortex-M chips include embedded code, DSP code, and a neural network model. All of that is folded into one software development flow in the Cortex-M52, giving programmers an easier development path that works with common machine learning frameworks and existing tools.
“The developer can code in a single language against a common API, achieve that performance uplift that they need in both DSP and ML elements of their application,” Williamson said. “There’s no need to understand the specific hardware details of the processor that sits beneath.”
Software developers need DSP and machine learning performance to take advantage of what AI can do. Previously, this meant using a CPU, DSP, and an NPU.
“That means that they would have to build the hardware and, once it was built, they may have to write, debug, and chain code across multiple chips or multiple processors within a single design,” he said. “I might need three separate toolchains, compilers, debuggers, and the developer would have to have that deep understanding of timing and memory access and scheduling of events across the multiple processors. A very complex task.”
Chip makers already have the Cortex-M52 and it will start appearing in silicon next year, he said.
Keeping an Eye on RISC-V
Other chip makers, from Intel, AMD, and Nvidia to a growing number of smaller companies, are aiming to bring more AI capabilities into IoT and embedded devices, but according to Enderle, Arm’s “potentially big exposure is from RISC-V which has been making inroads and the use of ASICs instead of processors for this.”
RISC-V hit the scene about a decade ago, giving chip makers and developers an open source chip design that RISC-V International, which is developing the chip architecture, can be used as an alternative to x86 and Arm. It’s a mantra that RISC-V International brought to its recent RISC-V Summit, pushing the idea that the architecture can be used in any device, from IoT and the cloud to PCs and data center servers.
It’s gaining some traction, with Apple putting controllers in its silicon and others like Meta, AMD, and Qualcomm are looking into the architecture.
Right now, the field is tipped in Arm’s favor. While there are about 10 billion RISC-V cores in the market, with more expected in the coming years, Arm estimates that there are 100 billion Cortex-M devices that have shipped.
However, RISC-V comes with the open vibe that developers and organizations have become accustomed to through Linux and other open software. Companies can license it, creating their own versions of the chip.
“It is cheaper, and the backers are more open,” Enderle said. “A lot of ARM developers have or are switching to this. Both Qualcomm and AMD are looking at this technology as well. “
The open, freely available nature of RISC-V also is being used in China, putting the technology into an ongoing controversy around processors between the country and the United States. U.S. lawmakers and the Biden Administration are considering restrictions on American companies that want to export RISC-V-based technology to China.
During a Q&A session with journalists, Williamson sidestepped a question about the Cortex-M52 possibly being a response to China’s growing interest in RISC-V, saying that Arm has had a “strong focus on bringing DSPs and machine learning performance out for several years now. This is the sort of next evolution in that roadmap to bring it to a lower performance and lower power consumption, more constrained devices. It’s been something that has been a long time and a key development focus for us.”
He also spoke about the significant advantage that Arm has in the partner and software ecosystem over RISC-V, pointing to the wide range of capabilities in the market for the Arm architecture — from the smallest embedded devices to large servers — all the while with consistent libraries and tools.
“The reason we have that really strong ecosystem of partners investing in our tools is they know that when they produce a tool, it can then be used across all of those products and across all of those different market areas where the technology could be deployed,” Williamson said. “It’s a really mutually valuable thing for us and our ecosystem. … That consistency is there to drive that scalability.”
However, while Enderle agreed that Arm is ahead with its ecosystem, RISC-V is still out there building out its capabilities and partnerships and is getting attention.
“Recent pricing actions and litigation against licensees have made developers leery of the platform and RISC-V is very similar without the baggage,” the analyst said.
