At Computex 2024, NVIDIA announced its new Vera Rubin GPU architecture. What started out as a talk on artificial intelligence changed when NVIDIA talked about updating its data center chip architecture.
NVIDIA’s new architecture will use newer HBM4 memory. The company officially announced that Rubin graphics processor will have 8 HBM4 memory stacks, and Rubin Ultra chips with 12 HBM4 stacks are planned for release later.
In addition, NVIDIA plans to introduce the successor to the Grace Hopper and Grace Blackwell superchips in the form of a new Vera Rubin board with a Vera processor (VR200), which combines GPU and CPU into a superchip. NVIDIA has not disclosed many specifications and details, but it is known that the platform will support new NVLink 6 connections with speeds up to 3600 GB/s and CX9 SuperNIC network connections up to 1600 Gbps.
NVIDIA has also announced the Blackwell Ultra chip, also known as the GB200 GPU with support for 12 HBM3E memory stacks, which will increase the supported memory from 192 GB to 244 GB. There was no information about graphics cards based on the Blackwell architecture, but the B100 and B200 are expected to be released this year.
NVIDIA announced that Rubin is scheduled to debut in early 2026. According to NVIDIA’s new one-year cycle, Ruby Ultra chips will appear in 2027. There was no direct mention of consumer graphics cards yet.
Sources: VideoCardz, Tom`s Hardware, Hardwareluxx