Computex Nvidia CEO Jensen Huang has shrugged off Big Tech's attempt to attack his networking strategy.

Speaking at an Nvidia event timed to coincide with Taiwan's Computex conference, Huang was asked for his reaction to the recent announcement of UALink, a GPU interconnect tech proposed and backed by AMD, Broadcom, Cisco, Google, HPE, Intel, Meta, and Microsoft.

UALink is all about connecting GPUs - the same job Nvidia gives to its NVLink technology.

At Computex, AMD CEO Lisa Su and Intel CEO Pat Gelsinger both used their keynote talks to promote UALink, and the forthcoming Ultra Ethernet effort to make the venerable protocol a better carrier of traffic generated by AI Applications. Both bosses emphasized standards-based networks as superior enablers of ecosystem-wide innovation.

Gelsinger even observed: "Customers don't want proprietary islands of infrastructure in their datacenters."

That's an interesting line of attack because Intel and AMD, and all enterprise-grade server makers, have a tricky relationship with Nvidia. On the surface it's all smiles, because nobody will risk not certifying their hardware to run Nvidia accelerators as doing so means missing out on the AI boom. Yet the entire x86 server ecosystem also wants to compete with Nvidia, making a push for open networking a way to challenge the GPU champ without threatening its core datacenter business.

Huang shrugged that off by pointing out that NVLink is already in its fifth generation, but UALink is just a proposal that will not compete with NVLink for years.

"By the time the first gen of UALink comes out, we will be at NVLink seven or eight ... We have some good ideas," he added, referring to future improvements to Nvidia's networking tech.

He wouldn't elaborate on those improvements, but earlier at the event announced that Nvidia's SpectrumX switching range has adopted an annual release cadence. "We will do something mechanically and electrically different every year," he said, describing Nvidia's release plan as "speed up, scale out."

That plan is already working. Huang said SpectrumX and the BlueField3 Super NIC released in late 2023 "will be a multibillion-dollar business inside a year."

The CEO concluded that networking is core business for Nvidia because the company has built an integrated stack that rivals can't match.

After Blackwell, Rubin

Speaking of that stack, Huang also used Nvidia's pre-Computex keynote to reveal a little about the successor to its Blackwell platform named "Rubin."

Rubin will feature a new Arm-based CPU called "Vera," a new GPU, and include "advanced networking with NVLink 6, CX9 SuperNIC and the X1600 converged InfiniBand/Ethernet switch."

Details? Sorry. Nvidia didn't offer any, other than to say an "Ultra" version of Rubin will follow a year after its first incarnation.

Other nuggets of news dropped by Huang included the debut of Nvidia Inference Microservices (NIMs), pre-configured and optimized container images containing an AI model and all the dependencies needed to get them running under Kubernetes. NIMs were also made free to members of the Nvidia Developer Program, meaning a $4,500 per GPU per year subscription to Nvidia's AI Enterprise suite won't be needed.

Additionally, Huang's Tuesday talk saw him completely deny alleged issues with Samsung's high-bandwidth memory that reportedly meant it would not be used in Nvidia accelerators.

"We just have to do the engineering but it's not done," he said. "There is no story there."

Riffing on Iron Man

That answer was provided during a media Q&A session for which Huang arrived 30 minutes late, and immediately interrupted to ask for his lunch on grounds he had missed that meal and his breakfast, meaning he had been forced to subsist on just "six Tic Tacs" for the entire day.

"CEOs aren't fed very well," he said.

Perhaps consciously, he appeared to offer a decent impersonation of Iron Man's alter ego Tony Stark as he took occasional bites of a turkey sandwich that he bemoaned was completely lacking mayonnaise, while uttering a few tangential words of reply to questions between bites but warning his lack of nourishment left him unable to handle multi-part questions as he was operating at 25 percent efficiency.

He ate while sitting on a stool that rotated beneath him, a phenomenon he attributed to the Coriolis effect, before getting serious about questions and declaring that the three "very dry slices of turkey" he managed to consume had revitalized him.

In response to a question about whether sufficient energy is available to realize the world's AI ambitions, he replied that Earth has a surfeit of solar energy and datacenters devoted to training models should therefore be located in sunny but unpleasant places where people don't want to live. Datacenters devoted to inferencing should, he suggested, be located closer to population centers, an arrangement he said will mean cities aren't deprived of energy for everyday activities.

Asked if he was concerned that China will develop AI hardware to rival Nvidia's own, Huang stated that he takes Chinese competitors "very seriously" and then called for another question ... one of which concerned Nvidia's AI PC ambitions and earned a remark that the RTX graphics card is his company's effort in the field and is present in machines from major manufacturers.

Huang's Sunday keynote earned him criticism in some quarters as it featured no gaming content, a perceived slight. His Tuesday Q&A therefore brought a question about whether he and Nvidia still have time for the gaming community.

The CEO explained that his keynote clocked in at two hours, leaving no time to talk gaming. But he reassured the Computex press corps that he adores gamers and those who chronicle the industry.

"Without PC gamers how could we have created the foundation for everything else?" he asked. ®

https://www.theregister.com//2024/06/04/nvidia_computex_nvlink_rubin/