On stage at the NVIDIA GTC 2025 conference in San Jose on Tuesday, the CEO of Nvidia, Jensen Huang, announced a multitude of new GPUs to descend the company’s pipeline in the coming months.
The most important thing was perhaps Vera Rubin. Vera Rubin, which should be published in the second half of 2026, will include dozens of memory teraoctets and a CPU designed by personalized Nvidia called Vera. Vera Rubin offers substantial performance elements compared to its predecessor, Grace Blackwell, claims Nvidia, in particular on the workloads of inference and the formation of AI.
When associated with Vera, Rubin can manage up to 50 Petaflops while making the inference, more than double the 20 Petaflops for the current blackwell fleas of the company, explains Nvidia. And Vera is about twice as fast at the CPU used in Grace Blackwell.
Rubin will be followed by Rubin Ultra in the second half of 2027, a collection of 576 GPU Vera Rubin, said Huang.
In the closer term – the second half of 2025 – NVIDIA will publish Blackwell Ultra, a group of Blackwell chips from Nvidia associated with dedicated CPUs. Blackwell Ultra will be available in several versions, including a containing 72 Blackwell chips.
On the distant horizon is Feynman. Huang gave few details on the architecture of Feynman, named after the American theoretical physicist Richard Feynman, except that he presents a vera processor. It is not known either when the first GPU Feynman will arrive.