Nvidia on Monday announced it would be offering its converged and artificial intelligence infrastructure to support not only x86 chips from Intel and AMD but also chips from ARM Ltd., the microprocessor intellectual property company that Nvidia is in the process of buying.
"ARM is a leading provider of 5G in mobile handsets, but as 5G RAN host infrastructure evolves, there is increasing demands to deliver the kinds of high performance, low power-per-watt that ARM CPUs are known for," said Ronnie Vasishta, head of telecom products at Nvidia.
"We believe that AI and machine learning is going to be an essential element of 5G networks, and AI and 5G are already essential elements of the applications that run on top of a 5G network," said Vasishta. Hence, there is a need to put AI acceleration next to network elements that run the radio access network, or RAN, stack, he said.
Nvidia had already announced in April, at its annual GTC conference, its partnerships to develop the 5G machines; the element of ARM-based chips is the new factor with Monday's press release.
The announcement comes on the opening day of the Mobile World Congress trade show in Barcelona, which was moved from its usual slot in February. It had been canceled in 2020 amidst the coronavirus pandemic.
The machines, which Nvidia refers to as AI-on-5G on a server, will consist of three parts: an Nvidia A100 GPU, a BlueField 2 data processing chip, or DPU, created from the assets that Nvidia acquired with Mellanox; and a processor, either ARM or x86. The machine runs Nvidia's Aerial A100 software stack for 5G networking and the various Nvidia AI libraries, such as CuDNN.
Nvidia has a roadmap that extends beyond the initial server configuration. The next step, coming sometime next year, is what Nvidia calls "AI-on-5G on a card." That brings A100 to a plug-in card with the next version of the DPU, BlueField 3, which will combine 16 embedded ARM "A-78" cores on the die, a "CPU cluster," as ARM puts it.
"This means you no longer need an external host," said Vasishta, what the company considers "self-hosted."
Next, sometime around 2024, Nvidia will offer a converged part containing GPU plus DPU plus ARM cores on a single die, the so-called BlueField 4 part. That chip is unrelated to the Grace ARM-based CPU that Nvidia disclosed in May.
The premise for combiningwith 5G is the renovation of society, from retail selling to industrial transpiration, said Vasishta.
"Every industry will be transformed in the next ten years because the forces of artificial intelligence and 5G connectivity are combining with digital automation to drive a revolution called the fourth industrial revolution."
Corporations want to use that combination of AI and 5G on open systems computing, including industry-standard servers, argued Vasishta.
"That's where the value is, the applications that will run on 5G."
Applications of the devices will initially lie in the domain of computer vision, Vasishta said in a media briefing.
"The first use cases are going to be using our Metropolis SDK, which is computer vision," said Vasishta. "Whether in a smart factory of retail, those elements of computer vision are becoming a lot more pervasive, and so are those 5G networks, because they need to move around, and so I think that would be one of the first [applications] you will see."
The first machines, coming later this year, are being built with partners including Mavenir, Radisys, Fujitsu, and Ericsson.
ARM technology has had a tough time cracking the data center, with only 1% of the world's data centers using ARM-based servers, Dion Harris, head of accelerated computing products at Nvidia.
To promote development on ARM, Nvidia is also offering a development kit that will include ARM CPUs from startup Ampere Computing Inc., and A100 GPUs, along with the BlueField 2 DPU.
Nvidia announced as well a collaboration with Google's cloud business unit to run an "innovation lab" to develop 5G applications. The company is in talks with other public cloud operators for similar initiatives, said Vasishta.