Arm Targets the AIoT with High-Performance Ethos-U85 NPU and Corstone-320 Platform
New accelerator boasts four times the peak performance and a 20 percent power efficiency boost over its predecessors.
Arm has announced its latest designs for the Artificial Intelligence of Things (AIoT) and edge artificial intelligence (edge AI): the new, faster Ethos-U85 neural processing unit (NPU) and a reference design platform, the Corstone-320, which combines it with an Arm Cortex-M85 CPU and a Mali-C55 image signal processor.
"Our family of Ethos-U NPUs were the world's first embedded AI accelerators and they're already in silicon from leading players," Arm's Paul Williamson told us during a pre-launch briefing on the company's latest designs.
"[The] Ethos-U85 delivers a 4x performance uplift for high-performance edge AI applications. Along with the 4x performance improvement, the U85 also delivers 20 percent higher power efficiency over the previous generation, is scalable from 128 to 2,048 MAC units, and delivers four TOPS [Tera-Operations Per Second at INT8 precision] at the highest performance configuration."
In addition to more raw power, the Ethos-U85 brings with it support for transformer networks. "It's a key feature," Williamson told us, "that allows faster customization and accelerates optimization for unique edge AI deployment. Transformer-based models can be adapted to different tasks more easily than convolutional networks [and] allow for more efficient hardware use, making it possible to deploy these models on edge devices with even quite limited compute resources."
The Ethos-U85 is backwards-compatible with the toolchain developed for earlier Ethos-U NPU models, Williamson has confirmed, while the Corstone-320 aims to deliver a quick-start platform — complete with virtual hardware support prior to silicon availability — by combining the NPU with the company's high-performance Cortex-M85 CPU core and a Mali-C55 image signal processor.
The Ethos-U85, though, won't be limited to the company's Cortex-M microcontroller-class parts: Williamson told us that it sees the NPU as suitable for pairing with its Cortex-A application-class cores too, to deliver acceleration for machine learning and artificial intelligence in robotics, industrial machine vision, and wearables.
Arm has not yet shared detailed benchmarks or power measurements for the Ethos-U85, which will vary by configuration, but Williamson has positioned it as "a milliwatts-level power envelope" design. The company has confirmed it is licensing the NPU to "early adopters" including Alif and Infineon now, with Williamson telling us that silicon is expected to land on the market some time in 2025.
More information on the Ethos-U85 is available on the Arm blog.