With Grace, Nvidia goes back to the future of data centers

jensen huang gtc21
Nvidia CEO Jensen Huang has big plans for the Grace CPU as the company prepares for a new assault on the data center market. (Screenshot)

Nvidia’s recently announced Grace CPU won’t be available for another two years, but that may feel like a short wait compared to how long it has taken Nvidia to get this far.

Grace is an Arm-based CPU that will use Nvidia’s NVLink CPU-GPU interconnection bridge to connect with Nvidia’s GPU to tackle massively large AI-driven workloads and supercomputing applications in the data center. It’s anticipated that Grace will run 10-times faster than Nvidia’s current X86-based data center systems.

Nvidia CEO Jensen Huang, during a press briefing this week, suggested Grace opens up a new market frontier for Nvidia. “My sense is it [the market for Grace] will be a very large new market just as GPUs were a zero billion dollar market,” he said. “That’s how we make a contribution to industry and how we invent the future.”

To make his point, he said the potential for Grace as a high-performance computing CPU for AI covers immense areas of study with data and processing. “A new scale of computer needs to be built that [uses] an Earthscale of data where a sensor is connected to everything on the planet to create a digital twin of Earth, to predict the weather down the square meter…and the physics of the geometry of Earth…We could do that for natural language understanding. Language is evolving continuously…There’s some very large data-driven sciences that need to be done with language models. Language is thought and thought is humanity’s ultimate technology."

He said natural language understanding with AI could cover multiple domains for how people talk in industries such as retail or insurance, and “we have to adapt models for every single one…so there could be 70 different languages and 100 different industries to train on data… forever.”

While AI may present a grand new frontier for Huang's ambitions, the data center servers in which the Grace CPU would operate represent a well-established market that Nvidia has been hovering around for years. The unveiling of Grace this week came a little over 10 years after Nvidia announced Project Denver, which itself had taken several years of work to advance to the announcement stage, according to published reports, and which at the time appeared to encompass a vision for Nvidia to build Arm-based CPUs for the data center to challenge the dominance of X86-based products there.

Grace also may prove to be just one part of Nvidia’s plan for a fresh assault on the data center market. The other part is Nvidia’s planned $40 billion acquisition of Arm. Analysts say if that deal is completed Nvidia gets more control over its data center destiny.

“If that deal closes, you could see this all in a way as Nvidia’s return to the data center,” said Shane Rau, research vice president, computing semiconductors at IDC. “It goes back to the Denver project, where they wanted to produce a server class of CPU. Now, with Grace, Nvidia will have the CPU, GPU and DPU, the three main sets for data center processing, in its portfolio. Arm is involved in silicon for all three of those, so Nvidia could own its data center ecosystem and the technology roadmap for that ecosystem. So, what Nvidia’s doing right now has them squarely focused on the data center.”

Kevin Krewell, principal analyst at Tirias Research, said in an email to Fierce Electronics, "Grace is one more proof point that Nvidia values the performance possible with the Arm architecture, and Nvidia's desire to push the Arm architecture to the highest-performance platforms."

Nvidia’s plans for Grace and its Arm deal also have the company squarely focused on a looming battle with Intel and AMD in the data center. Jack Gold, founder and principal analyst of J. Gold Associates, told Fierce Electronics by email, “Nvidia sees an opening against Intel and wants to be a full service provider for data center and cloud systems, particularly as it can leverage its GPU/DPU and AI strengths. But it won’t have an easy time getting rid of Intel (and AMD) x86 based systems, as it’s also true that a huge amount of code running in enterprise data centers and in the cloud is optimized for x86 (and not ARM), and it will be a huge task to convert that.”

During a press briefing about Grace, Nvidia acknowledged that X86 still had a place in the data center, and that Grace would begin by targeting the niche of largest-scale requirements being posed by increasingly large AI training models and high-performance supercomputing applications. Gold concurred that there are “some niches where Nvidia could very well compete and with its combined assets could prevail over Intel (although certainly Intel is not sitting still and is becoming more competitive in this space as well).”

He also cautioned, “Many vendors have tried to make ARM-based server chips a viable alternative to X86 in data centers and cloud deployments in the past (including AMD and Qualcomm) with very limited success (in fact both quietly exited that business fairly quickly after announcing products.) There are other companies trying to go the ARM route for servers, but so far it’s been a very niche marketplace with no real big winners.”

RELATED: Nvidia CEO Jensen Huang and the future of... everything