Nvidia Unveils Rubin and Rubin Ultra Data Center GPU Roadmap
Nvidia reveals its next-generation Rubin and Rubin Ultra data center GPUs at GTC 2025, promising massive performance leaps over the current Blackwell architecture.
Nvidia reveals its upcoming Rubin and Rubin Ultra data center GPUs at GTC 2025, outlining a clear roadmap for 2026 and 2027. Although the company currently focuses on bringing Blackwell B200 into full production and launching B300 later in 2025, CEO Jensen Huang already provides partners with detailed plans for the next generation of AI hardware.
The new Rubin platform introduces NVL144 and NVL576 configurations, addressing a naming correction where the current Blackwell B200 NVL72 actually contains 144 total GPU dies. The Rubin NVL144 rack remains drop-in compatible with existing Blackwell NVL72 infrastructure but delivers a massive 3.6 EFLOPS of dense FP4 compute, marking a 3.3X improvement over the B300.
Rubin also represents a major memory upgrade by transitioning from HBM3e to HBM4, with Rubin Ultra utilizing even faster HBM4e. While per-GPU memory capacity stays at 288GB, memory bandwidth increases significantly from 8 TB/s to 13 TB/s, ensuring the new architecture handles future AI workloads with much greater efficiency.