Nvidia shows roadmap with AI GPUs, Rubin architecture with HBM4 will be released in 2026 – Computer – News

Nvidia CEO Jensen Huang announced his new roadmap for data center products at Computex. The company plans to release a new architecture for AI GPUs called Rubin in 2026. This should follow the recently announced Blackwell GPUs.

During his Computex 2024 keynote, Huang live shared Nvidia’s data center lineup for the coming years. He initially announced that Nvidia will release a Blackwell refresh next year. That version supports twelve HBM3e stacks, up from eight currently. This Blackwell Ultra refresh should therefore support larger memory capacities compared to the first Blackwell B200 GPUwhich will hit the market later this year and offers 192GB of HBM3e memory.

The CEO also announced his plans for 2026 and 2027. Huang confirmed that Nvidia’s next generation of data center GPUs will be codenamed Rubin. The architecture is named after astronomer Vera Rubin. The manufacturer is not yet sharing many details about the new Rubin GPUs, as the current Blackwell cards are not even on the market yet. The manufacturer does confirm that the Rubin GPU will support faster HBM4 memory. This concerns eight HBM4 stacks. A year later, in 2027, Rubin Ultra will follow with twelve HBM4 memory stacks for greater total capacity.

Huang further confirmed the arrival of a new processor for data centers, which will succeed the company’s current Grace-Arm CPUs. Few details are known about these processors, which are codenamed Vera. There will also be a Vera Rubin superchip, in practice a PCB on which two Rubin GPUs are combined with a Vera processor, as the chip company previously did with its Grace Hopper superchip. That platform will also support new NVLink 6 switches, which should offer bandwidths of 3.6TB/s.

Nvidia does not share concrete release dates for the new products. There were previously rumors about the arrival of a Rubin GPU as a successor to Blackwell. Those rumors stated that the chips would go into production at the end of 2025. Analyst Ming-Chi Kuo reported that the GPUs are produced on TSMC’s N3 node, although Huang did not confirm this in his keynote on Sunday.