Arm Looks to Accelerate On-Device Intelligence for IoT Devices with Successive Announcements

In November 2023, Arm announced the acquisition of a minority stake in Raspberry Pi Ltd, extending the companies’ long-standing partnership as Arm strives to make inroads within the Internet of things (IoT) developer community.

Raspberry Pi has offered Arm-compatible Single-Board Computers (SBCs) since 2008, becoming a staple in computer science education by providing a foundational platform for simple input/output command-based control. More recently, Raspberry Pi has ramped up shipments for use in commercial and industrial markets, and its SBCs have proven well-equipped for developing high-performance IoT devices. Greater commercial usage is likely to have spurred Arm’s interest, as the company attempts to grow its market share for the compute platforms used in IoT devices.

Concurrently, Arm has responded to the growing calls within its developer community for on-device intelligence, extending its Cortex-M chip family portfolio to incorporate improved Machine Learning (ML) capabilities. Only 3 weeks after its agreement with Raspberry Pi, Arm announced the release of the Cortex-M52 processor, capable of accelerating advanced Artificial Intelligence (AI) workloads on the smallest IoT devices, such as smart sensors for industrial, medical, or wearable applications.

Under SoftBank’s ownership of Arm (after 2016), Masayoshi Son had envisioned a keystone role for Arm in the IoT ecosystem; however, Arm struggled to gain meaningful traction, epitomized by its largely unsuccessful acquisition of Treasure Data in 2018. Following Arm’s listing on Nasdaq in September 2023, these two latest announcements can be seen as a boost for Arm’s IoT strategy.

The release of Cortex-M52 addresses core engineering problems for the Artificial Intelligence of Things (AIoT)—ML performance, software flexibility and power consumption. To achieve the Cortex-M52 required ML performance level, a dedicated accelerator would be very low power, but would be too inflexible for varied AIoT workloads. A more complex chip with a separate Central Processing Unit (CPU), Neural Processing Unit (NPU), and Digital Signal Processor (DSP) would increase power consumption and complexity, with developers having to use multiple tool chains, debuggers, and compilers. Instead, the Cortex-M52 uses Arm’s Helium vector extensions to enable improved DSP and ML performance within the CPU itself.

Using this approach, the Cortex-M52 will enable an improved AI-powered audio and visual experience for applications, including wearables, automotive and industrial control, and predictive maintenance.

By tightening its relationship with Raspberry Pi, Arm strengthens its access to a significant pool of developers. To date, Raspberry Pi has sold more than 40 million SBCs worldwide, and since 2020, over 50% of its total sales came from industrial and commercial customers—a major shift from its original market in educational applications. In parallel, the release of Cortex-M52 provides a practical solution for next-generation SBCs in response to the increasing demand for on-device intelligence. Taken together, these developments should facilitate greater innovation and prototyping of AIoT devices over the next few years.

According to Arm, Helium extensions provide a significant improvement in DSP and ML performance compared with previous Arm Cortex-M family processors, enabling more compute-intensive ML inference algorithms in IoT devices without a dedicated NPU. The technology was previously integrated into higher performance CPU products used at the network edge, and by embedding it in the Cortex-M52, Arm says it can now bring this performance to more power-constrained IoT devices.

Processors such as the Cortex-M52 also have the potential to improve the security, privacy, and reliability of data processing, and to reduce overall system power use and latency by minimizing the amount of data transfer required to cloud-based servers. For all these reasons, IoT device developers should take advantage of the opportunity to efficiently incorporate ML inference algorithms in endpoint devices, and to enable more processing locally at the edge of the network, in order to differentiate their products in the emerging AIoT market.

Developers of IoT chips and modules should seek to start incorporating ML capability into their devices using technology such as that implemented in the Cortex-M52 to be ready for the growing requirement for AIoT, potentially upgrading from Arm’s previous generation Cortex-M33 and Cortex-M4 designs. For new entrants to these markets, open source and cheaper solutions based on RISC-V may provide an alternative approach, as discussed below.

Software and system developers already using Arm’s Cortex-M family for IoT devices should consider upgrading to devices based on the Cortex-M52, as a way to implement on-device intelligence.

For Raspberry Pi, the announcement not only cements its relationship with Arm, but also strengthens its access to developers within the Arm ecosystem. Arm’s commitment to integrating ML capabilities aligns well with emerging opportunities on Raspberry Pi. For example, a group of developers recently ran a Large Language Model (LLM) on a Raspberry Pi device and created their own AI chatbot server. Within the Arm Connected Community, Raspberry Pi should encourage adoption of its latest platform within the network of computer-science hobbyists and enthusiasts that are often at the forefront of innovation in the IoT devices space.

As the scope of AI endpoint intelligence grows, especially within commercial and industrial markets, Raspberry Pi will also have the flexibility to upgrade its SBC platform to the Cortex-M52 or subsequent cores, given the software compatibility maintained in the Arm roadmap.

The Emerging Battleground of AIoT for RISC-V and Arm

Finally, both strategic decisions from Arm are a sign of its awareness of the emerging threat of RISC-V in the IoT space. In Arm’s Initial Public Offering (IPO) prospectus, the company publicly recognized RISC-V architecture as a viable, royalty-free alternative to its proprietary blueprints. Many of Arm’s customers, including Qualcomm and indeed Raspberry Pi, have joined the RISC-V Foundation, a nonprofit dedicated to promoting the adoption of the RISC-V architecture and enablement of the RISC-V developer community. Establishing a stake in a strategic partner that has already flirted with RISC-V is an indication of Arm’s initiative in enacting a more direct influential position. An efficient new CPU design that should ease developers’ workloads to achieve the performance required for edge AI applications is a timely response, highlighting the performance and power advantages of Arm’s proprietary RISC architecture against its open-source counterpart, in a market segment where the lowest power consumption is vital.

While the acquired stake in Raspberry Pi and Cortex-M52 release should help Arm fend off the immediate threat imposed by RISC-V, Arm must also continue to be receptive to the needs of its developer ecosystem in the IoT space. By making it easier to implement more complex on-device intelligence at the endpoint and network edge, Arm can create more longer-term software lock-in for its architecture roadmap and in IoT markets where this has not historically been the case.