Arm Launches A New Generation Edge AI Accelerator Ethos-U85

On April 9, Arm announced the Ethos-U85 neural network processor (NPU), its MAC(multiplication-cumulation-computing) units can be expanded from 128 to 2048, and its computing power can achieve 4TOPS at 1GHz frequency, which can provide support for high-performance edge AI applications such as factory automation, commercial retail, and smart home. The press conference also launched a new iot reference design platform -Arm Corstone-320.

Today, microprocessors are deployed in many high-performance iot systems, including industrial machine vision, wearables, and consumer robots, and each scenario puts higher requirements on artificial intelligence capabilities. The Ethos-U85 was introduced with the goal of combining with the Armv9 Cortex-ACPU to accelerate the processing of machine learning (ML) tasks and provide more energy efficient edge reasoning capabilities to meet the higher demand for AI capabilities in these scenarios.

It is understood that Ethos-U85 has four times better performance and 20% better energy efficiency than its predecessor, and can support Transformer architecture and convolutional neural network (CNN) for AI inference, Transformer architecture is suitable for visual and generative AI use cases, and can be used to understand video, fill in missing parts of images. Analyzing data from multiple cameras for tasks such as image classification and object detection is very effective. The new Ethos-U85 NPU supports AI frameworks such as TensorFlow lite and PyTorch.

The Fthos-U85 provides developers with a chain consistent with the previous architecture, allowing partners to leverage existing investments for a seamless development experience. To date, the Arm EthOs NPU family has more than 20 licensed partners, with Alif Semiconductor and Infineon being early adopters of the new Arm Etho-U85 NPU.

The Arm corstone-320 iot Reference Design platform integrates Arm's highest performance Cortex-M cpus -Cortex-M85, Mali-c55 ISP and the new Ethos-U85 NPU. Deliver the required performance for a wide range of edge AI applications such as voice, audio and vision, such as real-time image classification and object recognition, or enable voice assistants with natural language translation capabilities on smart speakers. The reference design platform covers software, tools, and support, including Arm virtual hardware. The platform's combination of hardware and software will enable developers to start software development before the physical bibi is ready, accelerating product progress and reducing time to market for increasingly complex edge AI devices.

The first is the ability to provide more natural interaction. Traditional smart home, in order to use its intelligent functions, users need to download a new APP, which is not friendly to users. The natural language interaction capability of household appliances equipped with large models has been greatly improved. Using the reference design of Ethos-U85 and Corstone-320, small models compressed by large models supporting natural language interaction can be deployed to various household appliances more easily. To support users to interact directly with home appliances in natural language.

The second is that Arm can play the advantages of multi-field products in the cloud side. Large models are intelligent because of the large number of parameters. Most large models are currently deployed in the cloud. Home appliance manufacturers collect user data, etc., are concentrated in the cloud training and optimization. But at the same time, in order to enhance the user experience, the home appliance terminal side and the home gateway side will also deploy models supporting natural language response, which requires the device to achieve cloud-side collaboration. Ma Jian also added that there are strong products and unified tool chain support in the cloud, edge and end of the data center, so it has more ecological advantages.