ARM boosts machine learning for MCUs

ARM boosts machine learning for MCUs

Technology News |
By Peter Clarke

The Ethos-U55 is designed as a neural network accelerator for microcontrollers. The combination of Ethos-U55 and Cortex-M55 can provide a 480x improvement in machine learning performance, ARM reckons.

The Ethos-U55 joins the Ethos-N77, the Ethos-N57 and Ethos-N37 in the NPU family. Presumably ARM will also supply ‘P’ prefixed NPUs.

The Cortex-M55 is intended as a stand-alone microcontroller processor that can do all the conventional functions of an MCU but with enhanced support for processing artificial neural network maths and providing about a 15x performance boost. This comes from the inclusion of ARM’s Helium technology, which provides support for vector maths.

Arm Helium technology is the M-Profile Vector Extension (MVE) for the Arm Cortex-M processor series. Helium is an extension of the Armv8.1-M architecture and delivers a performance uplift for machine learning (ML) and DSP applications. The Cortex-M55 processor is the first Arm processor to support Helium.

Next: NPU

ARM Custom Instructions will be available to extend performance for specific tasks (see ARM enables custom instructions on Cortex-M).

The arrival of Cortex-M55 would extend the capability of Cortex-M processor cores from the sensor fusion and keyword detection they can do today into such areas as more generalized anomaly and object detection.


The Ethos-U55 is configurable with between 32 and 256 8×8 multiply-accumulate units capable of 32x uplift compared to the Cortex-M55. It is intended to be paired with a Cortex-M series processor core, such as the Cortex-M55 or others to provide yet higher performance appropriate to gesture detection, speech recognition and biometric awareness. This also comes with 25x more energy efficiency when performing these tasks.

To get higher machine learning performance than that ARM advocates use of one of the N prefixed Ethos cores.

The Ethos-U55 and Cortex-M55 are supported by a unified software development flow that integrates embedded code, DSP routines and neural networking code It supports machine learning frameworks such as Google’s TensorFlow and Facebook’s PyTorch, plus ARM’s own solutions.

Table: Ethos-U55 specifications. Source: ARM Ltd.

First silicon based on these cores is expected early in 2021 and will likely come in the form of microcontrollers and ASICs for or from Amazon, Alif Semiconductor, Bestechnic, Cypress, Dolby, Google, NXP, Samsung and STMicroelectronics.

Related links and articles:

News articles:

ARM adds Ethos cores to machine learning catalogue

ARM enables custom instructions on Cortex-M

If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News


Linked Articles