Arm Shows Off New NPU and IoT Reference Design at Embedded World

At Embedded World 2024, Arm announced two new IoT-targeted solutions to provide more AI computing power at the edge. As AI continues to gain momentum across industries, designers can use Arm's new Ethos-U NPU and IoT reference design platform to deploy power-efficient AI models to revolutionize edge computing.

 

The Ethos-U85 NPU can be integrated alongside Cortex-A or Cortex-M processors.

 

To learn more about Arm's new hardware, we spoke with Paul Williamson, senior VP and GM of Arm’s IoT business line, to find out how the new Ethos NPU and Corstone reference design will help designers bring more AI performance to the edge. 

 

Paul Williamson (left) discusses the new NPU and reference design with All About Circuit editor-in-chief Jeff Child (right) at Embedded World 2024 in Nuremberg, Germany. 

 

Efficient, Powerful Neural Processing

The Ethos-U85 neural processing unit (NPU) allows designers to add hardware-based AI acceleration to embedded devices. The Ethos-U85 expands Arm’s previous Ethos NPU portfolio, offering higher performance and support for transformer-based AI models. 

“This support is a key feature that allows faster customization and accelerates optimization for unique edge AI deployment,” Williamson said. “Ethos U85 offers not only weight by matrix for convolutional neural networks but also matrix by matrix multiplication, which is essential as a component for transformer networks. Because transformer-based models can be adapted to different tasks more easily than convolution traditional networks, transformer networks will drive new applications.”

 

The Ethos-U85 NPU improves upon previous generations by adding transformer support, lifting performance, and reducing energy consumption.

 

The Ethos-U85 NPU offers scalable performance levels from 256 GOPs to 4 TOPs at 1 GHz (128–2048 MACs). The U85 can also be integrated with existing Arm processors, including Cortex-A and Cortex-M series devices. Finally, the Ethos-U85 has continued to decrease energy consumption, reporting a 20% lower energy usage than past generations.

 

Embedded AI Reference Design

To showcase the Ethos-U85's performance and jump-start development, Arm has also announced its Corstone-320 IoT reference design. This dedicated platform includes supporting hardware, software, and development tools to help designers rapidly adopt the new Arm hardware and accelerate time to market.

 

The Corstone-320 reference design incorporates the latest Arm hardware and software tools.

 

The Corstone-320 has three key processing blocks: the Cortex-M85 processor, the Ethos-U85 NPU, and the Mali-C55 image signal processor to enable camera-connected applications. With these blocks working in tandem, designers can quickly develop solutions to complex AI-vision problems, including smart home or industrial applications.

“The reference design has been developed with real workloads in mind,” Williamson said. “Examples include battery-powered camera systems deployed in smart homes and low frame rate connected cameras that are being used in industrial production line and retail systems.”

 

Accelerating Edge AI Adoption

Machine learning and intelligence can extend beyond data centers to many embedded applications. These systems may see an uptick in AI deployments using the Ethos-U85 and Corstone-320 designs. 

 

Designers can use the entire Arm ecosystem to add the latest AI-focused hardware to their own embedded devices. 

 

“AI is everywhere. It's becoming more and more pervasive,” Williamson said. “As we see the rise of large language models and generative AI, it's clear that innovation can't slow down. It may seem the AI explosion came overnight, but the truth is, Arm has been preparing for this moment for a long time.”