Nvidia announces new, more powerful Vera Rubin chip made for AI | Nvidia
Nvidia CEO Jensen Huang said Monday that the company’s next-generation chips are “in full production” and will be able to deliver five times more AI computation than the company’s previous chips while delivering chatbots and other AI applications.
In a speech at the Consumer Electronics Show in Las Vegas, the leader of the world’s most valuable company revealed new details about chips coming later this year that Nvidia executives say are being tested by artificial intelligence firms in the company’s labs, at a time when Nvidia faces increasing competition from both rivals and its own customers.
The Vera Rubin platform, which consists of six separate Nvidia chips, is expected to launch later this year with the company’s flagship server featuring 72 graphics units and 36 new central processors.
Huang showed how they could be combined with more than 1,000 Rubin chips and arranged into “capsules,” and said they could increase the efficiency of producing items known as “tokens,” the basic unit of artificial intelligence systems, by 10 times.
But to achieve the new performance results, Huang said the Rubin chips use a special kind of data that the company hopes the broader industry will adopt.
“This is how we were able to achieve such a massive increase in performance, even though we only had 1.6 times the number of transistors,” Huang said.
While Nvidia still dominates the market when it comes to training AI models, it faces much more competition from traditional rivals like Advanced Micro Devices as well as customers like Alphabet’s Google when it comes to bringing the fruits of those models to hundreds of millions of chatbots and other technology users.
Much of Huang’s talk focused on how well the new chips would work for this task; This includes adding a new layer of storage technology called “context memory storage,” which aims to help chatbots provide quicker responses to long questions and conversations.
Nvidia also introduced next-generation network switches with a new type of connection called bundled optics. The technology, which is key to bringing together thousands of machines, competes with offerings from Broadcom and Cisco Systems.
In other announcements, Huang highlighted new software that could help driverless vehicles decide which path to take, leaving a paper trail that engineers can use later. Nvidia showed off research on software called Alpamayo late last year, and Huang said Monday it would be released more widely, along with the data used to train it so automakers can make evaluations.
“Not only do we open source the models, we also open source the data that we use to train those models, because that’s the only way you can truly trust how the models come out,” Huang said from a stage in Las Vegas. Last month, Nvidia bought talent and chip technology from startup Groq, whose executives include executives who helped Alphabet’s Google design its own AI chips.
Although Google is a major Nvidia customer, its own chips have emerged as one of Nvidia’s biggest threats, as Google has worked closely with Meta Platforms and others to chip away at the company’s lead.
In a question-and-answer session with financial analysts following his speech, Huang said the Groq deal “will not impact our core business” but could result in new products that expand the product line. At the same time, Nvidia is eager to show that its latest products can outperform older chips like the H200 that Donald Trump allowed into China.
The chip, a precursor to Nvidia’s current “Blackwell” chip, is in high demand in China, alarming China hawks across the U.S. political spectrum.
