While artificial intelligence algorithms seem to be everywhere, processing them on the most popular platforms requires powerful server GPUs in order to serve customers. Arm is launching a new specialized chip design that will deliver artificial intelligence acceleration capabilities in even the most affordable IoT devices starting next year.
According to the company, the ArmCortex-M52 is the smallest and most cost-effective processor designed for artificial intelligence acceleration applications. As Arm puts it, the British fabless company's latest design promises to deliver "enhanced" artificial intelligence capabilities to Internet of Things (IoT) devices without the need for a separate computing unit.
Paul Williamson, Arm's senior vice president and general manager of the Internet of Things business, emphasized that to fully realize the potential of artificial intelligence in the field of Internet of Things, machine learning optimization processing must be introduced into the "minimum and lowest power consumption" terminal equipment. Williamson pointed out that although artificial intelligence is ubiquitous, gaining "intelligence" from the vast amounts of data flowing through digital devices will require smarter and more powerful IoT devices.
The Cortex-M52 chip design uses Arm's Helium technology, adding 150 new scalar and vector instructions to the Armv8.1-MCortex-M series (including Cortex-M55 and Cortex-M85). Helium instructions improve the performance of machine learning algorithms by 5.6 times and digital signal processing (DSP) workloads by 2.7 times compared to the previous generation Cortex-M.
As Arm explained, security remains a crucial aspect. Cortex-M52 implements the latest security extensions of Armv8.1-M (PACBTI, ArmTrustZone). The new chip design also facilitates "modern development flows," according to Arm, giving developers access to the Cortex-M platform's unified tool chain and full support for artificial intelligence workflows.
Prior to the Cortex-M52, developers had to combine CPU, DSP and NPU units with three different software toolkits to achieve the ML and DSP performance natively provided by the new design.
Arm says that now just one tool chain is enough. Cortex-M52 is fully compatible with software written for Cortex-M55 and Cortex-M85, and the new chip will also enable pre-silicon software development through the Arm virtual hardware cloud platform.
Williamson said Cortex-M52 technology can be integrated into very low-cost IoT products through licensing, with chips costing $1-2 likely accounting for the majority of production volumes. The chip could also be integrated into "slightly feature-rich" IoT devices, he added.