Nvidia announces Rubin GPUs (2026), Rubin Ultra (2027), and adds Feynam to its multi-year GPU roadmap.

Nvidia announced updates to its data center roadmap for 2026 and 2027 at the company's GTC 2025 conference, showcasing the planned configurations for the upcoming Rubin and Rubin Ultra platforms, named after astronomer Vera Rubin. The company is already planning for transitions even as Blackwell B200 enters full production and Blackwell B300 is slated for the second half of 2025. During the announcement, CEO Jensen Huang noted an interesting historical point: "Blackwell was named wrong." He explained that Blackwell B200 actually has two dies per GPU, which changes the NVLink topology, meaning it would have been more appropriate to call the current solution NVL144 rather than NVL72—a naming convention Nvidia will adopt with the upcoming Rubin solutions.

The Rubin NVL144 rack will be drop-in compatible with existing Blackwell NVL72 infrastructure and will contain 144 total GPU dies. Compared to the Blackwell Ultra B300 NVL72, which offers 1.1 EFLOPS of dense FP4 compute, Rubin NVL144 will deliver 3.6 EFLOPS of dense FP4 compute. Rubin will also provide 1.2 ExaFLOPS of FP8 training, compared to only 0.36 ExaFLOPS for B300, representing a 3.3X overall improvement in compute performance. The platform will mark a shift from HBM3/HBM3e to HBM4 memory. While memory capacity remains at 288GB per GPU, the same as B300, bandwidth will improve from 8 TB/s to 13 TB/s, and a faster NVLink will double throughput to 260 TB/s total. Additionally, a new CX9 link between racks will provide 28.8 TB/s, double the bandwidth of B300's CX8. The Rubin platform will also introduce the Vera CPU, replacing current Grace CPUs, featuring 88 custom ARM cores and 176 threads with a 1.8 TB/s NVLink core-to-core interface to link with the Rubin GPUs.

Rubin Ultra, scheduled to land in the second half of 2027, will feature the same Vera CPU but with significantly enhanced GPU capabilities in a new NVL576 configuration. This platform will pack up to 576 GPUs in a single rack with four GPU dies per package to boost compute density. The inference compute with FP4 will reach 15 ExaFLOPS, with 5 ExaFLOPS of FP8 training compute—approximately 4X the compute of Rubin NVL144. Where NVL144 Rubin offers 75TB total of "fast memory" per rack, Rubin Ultra NVL576 will provide 365TB. The GPUs will feature HBM4e memory with 4.6 PB/s of bandwidth, complemented by 1TB of HBM4e per four reticle-sized GPUs delivering 100 PetaFLOPS of FP4 compute. The NVLink7 interface will be 6X faster than Rubin's with 1.5 PB/s of throughput, while CX9 interlinks will see a 4X improvement to 115.2 TB/s between racks. Looking further ahead, Nvidia's next data center architecture after Rubin will be named after theoretical physicist Richard Feynman.