Teaming Up with Arm, NXP Ups Its Place in the Machine Learning Industry

October 19, 2020 by Jake Hertz

This week, NXP Semiconductors announced two strategic partnerships that help solidify its place in the ML industry. We interviewed experts at NXP to get the inside scoop.

One of the most popular topics in the technology industry, even for electrical engineers, is machine learning. The newest company to make headlines in the field is NXP Semiconductors with two big announcements today. 

Looking to further establish its place in the machine learning industry, NXP has made two strategic partnerships, one with Arm and one with Canadia-based Au-Zone. All About Circuits had a sit down with executives at NXP to understand what the news really means.


NXP and Arm Develop New NPU

On the hardware side of things, NXP announced today that it has been collaborating with Arm as the lead technology partner on the new Arm Ethos-U65 microNPU (neural processing unit). This technology partnership allows NXP to integrate the Ethos-U65 microNPU into its next generation of i.MX applications processors with the hopes of delivering energy-efficient, cost-effective ML solutions. 

NXP is particularly excited about this partnership because this new microNPU is able to maintain the MCU-class power efficiency of the Ethos-U55, but is capable of being used in systems with higher performance Cortex-A-based SoCs.


Block diagram of the Ethos-U55 NPU

Block diagram of the Ethos-U55 NPU. Image used courtesy of Arm

Some standout features of the Ethos-U65 include model compression, on-the-fly weight decompression, and optimization strategies for DRAM and SRAM.


i.MX simplified system block diagram

i.MX simplified system block diagram. Image used courtesy of NXP


What’s particularly unique about this SoC is that the NPU works alongside a Cortex-M based processor. In our interview, Ben Eckermann, Senior Principal Engineer and Systems Architect at NXP Semiconductors, explained why this feature is advantageous.

Eckermann explains, “What's key here is that, similar to the U-55, [the Ethos-U65] doesn't attempt to do everything as one standalone black box. It relies on the Cortex-M processor sitting beside it."

He continues, "The Cortex-M processor is able to handle any network operators that either occur so infrequently that there's no point in dedicating hardware resources in the U-65 to it or some that just don't provide you enough bang for your buck, where some things can be done efficiently on the CPU like the very last layers of a NN.” 


NXP and Au-Zone Create New ML Software

On the software side of things, NXP today announced that it has established an exclusive partnership with Au-Zone to expand NXP’s eIQ machine learning (ML) software development environment. 


The eIQ development environment

The eIQ development environment. Image used courtesy of NXP 

What NXP was really after was Au-Zone’s DeepView ML Tool Suite, which is said to augment eIQ with an intuitive, graphical user interface (GUI) and workflow. The hope is that this added functionality will make the development, training, and deployment of NN models and ML workloads straightforward and easy for designers of all experience levels.

The tool includes features to prune, quantize, validate, and deploy public or proprietary NN models on NXP devices.


DeepView workflow

DeepView workflow. Image used courtesy of NXP


Together, Au-Zone and NXP look to optimize NNs for NXP-based SoCs, providing developers with run-time insights on NN model architectures, system parameters, and run-time performance.

A key feature of this run-time inference engine is that it optimizes the system memory usage and data movement uniquely for each SoC architecture.

Gowri Chindalore, head of NXP's business and technology strategy for edge processing, claims that this feature offers customers a “double optimization," optimizing both the neural network and then further optimizing for the specific hardware. 


Making NNs More Accessible

With the introduction of the Arm Ethos U-65 microNPU, NXP will be able to provide new functionality and energy savings in future lines of i.MX application processors. This may make way for more powerful and low-energy designs for IoT and other edge applications. 

Introducing Au-Zone’s DeepView Tool Suite will also benefit design engineers because the training, optimization, and deployment of NNs will not only be made more simple but will also be optimized for the specific hardware they are running on.

This too should only benefit future developments in IoT and edge applications on NXP-based SoCs.