eWEEK content material and product suggestions are editorially unbiased. We could earn money whenever you click on on hyperlinks to our companions. Learn More.
NVIDIA CEO and cofounder Jensen Huang took to the stage on the NVIDIA GTC convention in San Jose on March 18 to point out the upcoming roadmap for the corporate, which has thrived over the previous couple of years by offering infrastructure for generative AI.
“Every single layer of computing has been transformed by the arrival of generative AI,” Huang stated.
NVIDIA introduced two new chips for working AI, a unifying working system for “AI factory” knowledge facilities, and a reasoning platform for robotics.
“Where we had a retrieval computing model, we now have a generative computing model,” he stated.
Vera Rubin and Blackwell Ultra are NVIDIA’s subsequent chips
NVIDIA’s next-generation GPU structure shall be Vera Rubin, named after the astronomer who found darkish matter. Vera Rubin shall be accessible within the second half of 2026, and can present 3.Three occasions the efficiency of a Blackwell Ultra in an analogous rack setup. Everything is new besides the chassis, Huang stated Vera Rubin shall be adopted within the second half of 2027 by an enhanced model: Rubin Ultra. NVIDIA acknowledged it presents 14x the efficiency of Blackwell Ultra.
NVIDIA will “gracefully glide” into manufacturing of the tuned model of the Blackwell chip later this 12 months, Huang stated. Blackwell Ultra can have 1.5 extra FLOPs, 1.5 occasions extra reminiscence, and two occasions extra bandwidth than the unique Blackwell.
The next-generation NVIDIA chips shall be named Feynman, after the physicist Richard Feynman.
Blackwell Ultra AI Factory constructed out with Blackwell Ultra
In order to supply the spine for the huge variety of generative AI workloads, together with agentic AI, being processed in knowledge facilities throughout the globe at the moment, NVIDIA unified a number of applied sciences, together with Blackwell Ultra, into its AI manufacturing unit platform. It contains the Blackwell Ultra with the NVIDIA GB300 NVL72 rack-scale resolution and HGX B300 NVL16. Inside, 72 Blackwell Ultra GPUs and 36 ArmNeoverse-based NVIDIA Grace CPUs are linked collectively. It additionally hyperlinks collectively the NVIDIA Spectrum-X Ethernet networking platform and NVIDIA Quantum-X800 InfiniBand platform.
The Blackwell Ultra AI manufacturing unit shall be accessible after the Blackwell Ultra chip is within the second half of 2025.
Huang framed NVIDIA’s AI infrastructure tasks as “scaling up” to “scale out,” permitting for nice effectivity and subsequently better income for its prospects. NVIDIA partnered with many of the main cloud suppliers, together with Microsoft Azure and Google, to develop the platform.
NVIDIA additionally introduced a Quantum-X photonics swap, which has the Spectrum-X Photonics Ethernet and Quantum-X Photonics InfiniBand packaged inside. Quantum-X built-in silicon is coming within the second half of 2026.
Introducing Dynamo, the “operating system” of the AI manufacturing unit
Another side of the Blackwell AI manufacturing unit shall be Dynamo, open-source inference software program that Huang known as the “operating system” of the AI manufacturing unit idea. It replaces the NVIDIA Triton Inference Server for orchestrating inference for reasoning AI and maximizing income as measured in opposition to the price of producing tokens. (Tokens are the chunks of data generative AI places collectively to make content material, equivalent to symbols, letters, or phrases.) It can be utilized to speed up the usage of AI inference at AWS, Meta, Microsoft Azure, and different main cloud platforms.
GROOT N1 is a basis mannequin that teaches robots to stroll
Huang additionally launched Isaac Groot N1, a…
