At GTC 2025, Nvidia has shared its roadmap for data centre GPUs, revealing a series of advancements promising to reshape the AI and HPC landscapes. The company's vision extends beyond the immediate future, encompassing refinements to its Blackwell architecture, introducing the Rubin (and Rubin Ultra) platform and unveiling a brand-new architecture codenamed Feynman.

The Blackwell Ultra NVL72, slated for release in the latter half of this year, marks the first step in this evolution. This system will use 72x Blackwell Ultra chips, each featuring GPUs and 288 GB of HBM3e memory, delivering a claimed 1.5x performance boost over the existing GB200 NVL72 platform. The complete rack, which also includes 36 Arm Neoverse-based Nvidia Grace CPUs, will boast 20 TB of HBM memory and 40 TB of fast memory.

Images via Nvidia

Looking further ahead, Nvidia has detailed its Rubin and Rubin Ultra platforms. Rubin, the first architecture built on Nvidia's next-generation design principles, is expected to debut in the second half of 2026. The company is developing the Rubin NVL144 platform, which will house Rubin chips featuring two GPUs and 288 GB of HBM4 memory. This platform will also pack 75 TB of fast memory Vera CPUs, each with 88x Custom ARM cores with dual threads. This whole suite promises to deliver up to 3.3x the performance of the Blackwell Ultra NVL72 platform.

Built on top of the Rubin NVL144, Nvidia plans to launch the Rubin Ultra NVL 576 with 365 TB of fast memory in H2 2027. This platform will pack the same Vera CPUs as the Vera Rubin NVL144, but instead of the standard Rubin chips, it will use Rubin Ultra, featuring four GPUs and 1 TB of HBM4e memory. As such, one can expect up to 14x more performance than the Blackwell Ultra NVL72 platform.

Lastly, Nvidia has provided a glimpse into its long-term vision with the Feynman architecture, which is anticipated to arrive in 2028. Details are still scarce, but we learnt it will leverage next-generation HBM memory technology.

KitGuru says: Do you think Nvidia will use any of these new architectures for products other than those aimed at HPC and the data-centre segments?

Become a Patron!