Ethos-U85 Adds Transformer Support

Arm’s EthosU85 makes Transformer-based models available for Ethos users implementing edge inference. Transformers are rapidly becoming a dominant deep-learning technology, moving from the datacenter into edge applications. The company’s new intellectual property (IP) also quadruples the performance available compared with the prior generation.

The updated neural processing unit (NPU) increases its MAC count from 512 to 2,048 while improving power efficiency by 20%, according to Arm. Dedicated hardware accelerates the math required for Transformers.

Edge processing reduces latency by eliminating round trips to the cloud; communication-bandwidth requirements also drop, and data privacy improves. The new capabilities brought about by Transformers are now desirable at the edge as they can provide better responses to text, voice, and visual events. The U85’s additional processing capability compared with prior generations will serve more demanding applications such as factory equipment.

Such equipment is typically line-powered, but power is always a consideration at the edge, whether running off batteries or the mains, and in the datacenter. However, Arm has declined to publicize its power and power-efficiency numbers in a manner that can be compared to other NPU IP.

Available for licensing now, the U85 accepts models designed in the popular TensorFlow and PyTorch platforms. It supports hosting by CortexA CPUs with the Armv9 architecture in addition to select CortexM, CortexR, and Neoverse CPUs.