Nvidia is breaking into the server CPU market with a new chip built on the ARM architecture.
The CPU, codenamed Grace, won’t arrive until early 2023. But according to Nvidia, the chip will deliver 10 times the performance over the leading x86 server chips when it comes to AI and large-scale data science workloads.
The product may pose a competitive threat to Intel, which has long dominated the server chip market. For years now, Nvidia has been supplying the GPUs used in data centers and supercomputers to power 3D simulations. The Grace chip helps fill out the company’s catalog, enabling server customers to buy both CPUs and GPUs from Nvidia.
“Nvidia is introducing Grace as the volume of data and size of AI models are growing exponentially,” the company said. “Today’s largest AI models include billions of parameters and are doubling every two-and-a-half months. Training them requires a new CPU that can be tightly coupled with a GPU to eliminate system bottlenecks.”
Grace debuted at Nvidia's GTC event, where the company talked up its latest advancements in enterprise technologies. The chip stands out by enabling 900GB/s connection between the Grace CPU and Nvidia GPUs for a 30 times higher aggregate bandwidth over the leading server chip system. The CPU also uses an LPDDR5x memory system to offer better bandwidth and energy efficiency over DDR4 memory, according to Nvidia.
However, the company points out the Grace CPU is a “highly specialized processor,” designed to serve niche markets. So it’s probably ill-suited for general purpose workloads.
Nevertheless, the chip is already attracting customers. Nvidia says the Swiss National Supercomputing Centre and the US Department of Energy’s Los Alamos National Laboratory both plan on building supercomputers around the Grace CPU. Expect them to go online in 2023.
The Grace CPU also uses the computing architecture from ARM, a company that Nvidia is trying to acquire if government regulators will let it. ARM chips are best known as low-energy processors used to power smartphones. But over the years, the CPU architecture has also been gaining traction in data centers and supercomputers as an energy-efficient alternative to x86 server chips.
Nvidia CEO Jensen Huang said in an investors call that he’s hoping his company can close the acquisition with ARM in 2022. However, he added that Nvidia's technology will continue supporting both ARM and x86 server chips.
It was not mentioned if whether Nvidia's Grace CPU will ever target PCs. But in 2011, the company announced Project Denver, an attempt to create ARM-based chips for PCs, servers, and supercomputers.