Arm Goes Heavy on Machine Learning with New Ethos Family of Neural Processing Units

Arm is continuing to push machine learning acceleration onto edge devices, announcing the launch of two new core intellectual properties (IPs) in its neural network processor range: the Arm Ethos-N37 and Ethos-N57, along with the Mali-G57 graphics processor and Mali-D37 display processor.

Where deep learning was once the province of supercomputers, the march of computational technology is increasingly pushing it towards the edge - even into battery-powered devices, with the latest smartphones using artificial intelligence for everything from local voice recognition to image enhancement for camera input. It's into this market Arm is looking to carve a niche, following up its earlier machine learning accelerator launch with two new cores: the Ethos-N37 and Ethos-N57.

"AI intelligence is now everywhere from gaming devices to digital TVs (DTVs), and increased compute at the endpoint is required to allow for these responsive experiences," claims Arm's Paul Williamson of the launch. "For example, intelligent experiences in DTVs range from smart assistant voice commands, real-time translation for shows in another language, and face recognition to enhance parental controls. To deliver this, I’m pleased to announce that Arm is launching two new mainstream ML processors, as well as our latest Mali graphics and display processors. Together, this IP represents Arm’s ability to scale, bringing premium experiences to everyday and ultra-efficient consumer devices."

The Ethos neural processing units (NPUs) are designed to offer the same technology as Arm's earlier Arm Machine Learning (Arm ML) processor, now rebranded to the range-topping Ethos-N77, with smaller power and footprint requirements. Where the Ethos-N77 offers up to four trillion operations per second (4 TOP/s) with 1-4MB of internal memory, the Ethos-N57 offers up to 2 TOP/s with 512kB of memory and the Ethos-N37 still less at up to 1 TOP/s with the same 512kB of memory. All are optimised for Int8 and Int16 workloads, with Arm claiming an "over 200 percent performance uplift" but refusing to specify to which competing chip it's comparing.

At the same time, Arm has announced the Mali-G57 graphics processor with a claimed 1.3x performance increase over the Mali-G52 along with a similar increase in energy efficiency. Here, too, the company is pushing machine learning workloads, claiming that the design offers a 60 percent improvement in machine learning performance compared to the previous generation. The company is also offering the Mali-D37, claimed to be the most area-efficient display IP the company has yet released - as little as 1mm² on a 16nm process while still offering 2k resolution support.

The company has published blog posts detailing the Ethos NPU, Mali-G57, and Mali-D37 products, but has not yet indicated when its customers will begin shipping the IP in development boards or commercial products.