At the Nvidia GTC 2025 conference, company CEO Jensen Huang unveiled a series of new GPU products scheduled for release in the coming months.
Among these, the most notable is the Vera Rubin GPU. Set to launch in the second half of 2026, this GPU will feature dozens of gigabytes of memory and incorporate Nvidia's self-designed Vera CPU. According to Nvidia, Vera Rubin will deliver significantly enhanced performance in AI inference and training tasks compared to its predecessor, Grace Blackwell.
Vera Rubin is actually a dual-GPU product, capable of reaching 50PFLOPS in AI inference performance—more than double the 20PFLOPS achieved by the current Blackwell chip. Additionally, the Vera CPU's processing speed is approximately twice that of the CPU in Blackwell.
Following Vera Rubin, Nvidia plans to introduce Rubin Ultra in the second half of 2027. This product will integrate four GPUs into a single package and achieve a performance of 100PFLOPS.
In the nearer term, Nvidia will release the Blackwell Ultra GPU in the second half of 2025. This product will be available in multiple configurations, with each chip delivering the same 20PFLOPS AI performance as Blackwell but increasing memory capacity from 192GB to 288GB.
Additionally, during his keynote presentation, Huang briefly introduced the Feynman GPU. Named after American theoretical physicist Richard Feynman, specific details about its architecture have not yet been disclosed, but it is known to include the Vera CPU. Nvidia plans to launch the Feynman GPU in 2028 as the successor to Vera Rubin.
