Arm Pairs Low-Power Processors with 5G Phones for On-Device AI


A new processor from Arm Holdings is enabling original equipment manufacturers produce connected devices with neural networks that mimic thought processes in the human brain.

The development has implications for 5G wireless and Internet of Things devices.

The processor from the British-based semiconductor designer promises to help enterprises cut costs for 5G smartphones and Internet of Things equipment by reducing reliance on the cloud with end-point artificial intelligenceOpens a new window .

New IPOpens a new window adds an integrated neural processor unit – the Ethos NPU – to the Cortex-M microarchitecture that Arm unveiled last year. The combination accelerates the modeling calculations that drive machine-learning algorithms.

The low-power design of Arm’s Cortex-M55 processor uses less real estate than systems-on-chips (SoCs), which combine central processor units with graphics engines. This means the tandem can be fitted more easily into sensors, wearables and smartphones. At the same time, the processor’s networking capability lowers latency by reducing the distances that sound, images and video must travel for parsing and instruction.

The result is an architecture that boosts digital signal processing for applications that are deployed at the network edge. And it bypasses the need for more of the centralized compute and storage infrastructureOpens a new window that enterprises must build themselves or rent from third-party providers.

5G Focused

Already boasting 160 billion chip deployments since its founding nearly three decades ago, Arm’s Cortex-M instruction-set architectures (ISAs) anticipate the exponential rise in connected devices that will result from the global rollout of 5G that began last year.

The next-generation telecom technology’s expanded frequency spectrum increases bandwidth for voice and gesture recognition, object recognition and classification, biometric awareness and facial recognition. It also underpins everything from streaming video to self-driving cars.

The Cortex-M familyOpens a new window of processors is designed to handle such tasks thanks to a microcontroller that users can program for intended applications. The microcontroller’s reduced ISA lowers energy consumption for processes executed over the design’s 32-bit cores.

Helium Filled

The Cortex-M55 ISA supports “HeliumOpens a new window ,” a vector extension that processes data and instructions simultaneously. With Helium, the low-power systems at work in sensors and control devices can better parse digital images and voice commands.

Developed for data centers, Helium increases the speed and accuracy of digital signal processing. It uses so-called Single Instruction Multiple Data methodology (SIMD) to combine instructions into a single operation.

A dedicated toolkit aids developers in tailoring the ISA’s floating-point registers for applications-specific processes. The Cortex-M55 also is augmented with performance-monitoring, debugging and security enhancements for IoT.

Doing the Math

Integrating the Ethos-U55 accelerates machine-learning calculations over the Cortex-M55 ISA. It spells the microcontroller by handling the algorithmic operations, with the company claiming a performance boost of almost 500% over Cortex-M’s running the numbers solo.

The microNPUOpens a new window can unpack more lightly-configured training models for inferencing over 8 cores for low-precision tasks, such as signal commands and motion detection. The Ethos-U55 scales up to 16 cores for operations like image recognition from stills and video that requires higher degrees of precision.

To reduce memory and allied energy consumption, the microNPUs work offline to optimize the neural networks they support. These include convoluted networks that feed data forward for image recognition and recurrent networks that reuse information for predictive analytics.

Licensed to Build

Arm says that on-device execution of machine-learning algorithms delivers the end-point artificial intelligence that renders relevant a full spectrum of data unlocked by 5G. That’s because shuttling to and from inference engines in the cloud costs milliseconds that can erode its value.

While Arm, a subsidiary of Japan’s Softbank, makes no siliconOpens a new window , licensees are working to create chips and SoCs that weave the low-power designs into the processors that original equipment manufacturers (OEMs) embed in their 5G-ready devices. These should begin arriving next year.

Arm cites a studyOpens a new window by McKinsey management consultants that points to increases in hardware spending at the network edge. With end-point AI the catalyst for real-time decision-making, part of the revenue redistribution that is foreseen as corporations adopt the new technology will come at the expense of cloud platforms.