On stage at NVIDIA’s GTC 2025 conference in San Jose on Tuesday, CEO Jensen Huang announced many new GPUs that will be appearing in the company’s product pipeline over the coming months.
Perhaps the most important thing is Vera Rubin. The Vera Rubin, scheduled to be released in the second half of 2026, will feature dozens of gigabytes of memory and a custom Nvidia-designed CPU called Vera. Vera Rubin offers a substantial performance uplift, especially when it comes to AI guessing and training tasks, compared to its predecessor, Grace Blackwell.
When combined with Vera, Rubin (technically two GPUs) can manage up to 50 petaflops while performing inference (i.e. running AI models). Plus, Vera is about twice as fast as the CPU used by Nvidia’s Grace Blackwell GPUs.
Rubin will be followed by Rubin Ultra in the second half of 2027. This is a collection of 4 GPUs in a single package that offers performance of up to 100 petaflops.
Nearby Horizon – H2 2025 – Nvidia releases Blackwell Ultra, a GPU that comes in several configurations. The single ultra chip offers the same 20 petaflops with the same AI performance as Blackwell, but has 288GB of memory. It will rise from 192GB of Vanilla Blackwell.
On the far horizon is the Feynman GPU. During his keynote speech, Huang gave little details about Feynman’s architecture, named after American theoretical physicist Richard Feynman. Nvidia plans to launch Feynman to the market in 2028, taking over Rubin Vera.
Updated 3/18 3:07 PM Pacific: Previous versions of this story showed Vera Rubin had “dozens of terabytes” of memory. In fact, it has “dozens of gigabytes” of memory. I regret the error.
Source link
