Semiconductor and software design company Arm is doubling down on edge AI hardware, a market that’s expected to be worth $1.15 billion by 2023. It today announced two new AI-capable processors — the Arm Cortex-M55 and Ethos-U55, a neural processing unit (NPU) — designed for internet of things (IoT) endpoint devices, alongside supporting software libraries, toolchains, and models. The company claims that the two chips, which are expected to arrive in market in early 2021, together will deliver an uplift of up to 480 times in machine learning performance in certain voice and vision scenarios.

“[Machine learning] processing on low-power endpoint devices is critical to realizing the full potential of AI for IoT,” wrote the company in press materials. “An extended range of advanced hardware capabilities is required to enable innovation and scale.”

The Cortex-M55, the newest member of Arm’s Cortex-M processor portfolio of cost-optimized and power-efficient microcontroller devices, delivers up to a 15 times uplift in AI performance on its own compared with previous Cortex-M generations, as well as custom instructions and configuration options. Like the Ethos-U55, it’s available in a reference design — Corstone-300 — that ships with a number of secure subsystems and a toolkit with which to build secure embedded systems.

The Cortex-M55 also has the distinction of being the first system-on-chip based on Arm’s Helium technology, an extension of the Armv8.1-M architecture that is optimized for low-power silicon and adds over 150 new scalar and vector instructions. The integer Helium enables efficient compute of 8-bit, 16-bit, and 32-bit fixed-point data, the last two of which are widely used in traditional signal processing applications such as audio processing. As for the 8-bit fixed-point format, it’s common in machine learning processing such as neural network computation and image processing, complementing floating point data types including single-precision floats (32-bit) and half-precision floats (16-bit).

According to Arm, Helium enables a performance boost of up to 5 times for digital signal processing and 15 times for machine learning with the Cortex-M55. It additionally allows for advanced memory interfaces to provide speedy access to machine learning data, and it builds in Arm’s TrustZone systemwide embedded security technology.

Ethos-U55 — which is intended to be paired with a Cortex-M processor like the Cortex-M55, Cortex-M33, Cortex-M7, or Cortex-M4  — packs between 32 and 256 configurable compute units capable of achieving up to 32 times machine learning performance uplift compared with the Cortex-M55. Pitted against the Cortex-M7, Arm claims that the Ethos-U55 and Cortex-M55 are up to 50 times faster in terms of speed to inference and up to 25 more energy efficient in tasks like voice activity detection, noise cancellation, two-mic beamforming, echo cancellation, equalizing, mixing, keyword spotting, and automatic speech recognition.

On the software side, both the Ethos-U55 and Cortex-M55 benefit from a unified software development flow that folds embedded code, digital signal processor code, and AI model code into one. Importantly, it plays nicely with popular machine learning frameworks like Google’s TensorFlow and Facebook’s PyTorch, plus Arm’s own solutions.

The unveiling of the Ethos-U55 comes after Arm took the wraps off of the Ethos-N57 and Ethos-N37, which both feature voice recognition and always-on capabilities. (Ethos, which launched recently, is a product suite focused on solving complex AI compute challenges with sensitivity to battery life and cost.) Around the same time in October, the company revealed its roadmap for Mbed OS, the embedded software operating system that’s meant to serve as a foundation for smart and connected devices.


The audio problem: Learn how new cloud-based API solutions are solving imperfect, frustrating audio in video conferences. Access here