Onstage at Nvidia’s GTC 2025 conference in San Jose on Tuesday, CEO Jensen Huang introduced a slew of latest GPUs coming down the corporate’s product pipeline over the subsequent few months.
Maybe essentially the most important is Vera Rubin. Vera Rubin, which is ready to be launched within the second half of 2026, will function tens of gigabytes of reminiscence and a customized Nvidia-designed CPU referred to as Vera. Vera Rubin delivers substantial efficiency uplifts in comparison with its predecessor, Grace Blackwell, Nvidia claims, significantly on AI inferencing and coaching duties.
When paired with Vera, Rubin — which is 2 GPUs in a single, technically — can handle as much as 50 petaflops whereas doing inference (i.e., operating AI fashions), greater than double the 20 petaflops for Nvidia’s present Blackwell chips. Furthermore, Vera is about twice as quick because the CPU utilized in Nvidia’s Grace Blackwell GPU.
Rubin will likely be adopted by Rubin Extremely within the second half of 2027, a set of 4 GPUs in a single bundle delivering as much as 100 petaflops of efficiency.
On the close to horizon — H2 2025 — Nvidia will launch Blackwell Extremely, a GPU that’ll are available a number of configurations. A single Extremely chip will provide the identical 20 petaflops of AI efficiency as Blackwell, however with 288GB of reminiscence — up from 192GB in vanilla Blackwell.
On the far horizon are Feynman GPUs. Huang through the keynote gave few particulars about Feynman’s structure, named after American theoretical physicist Richard Feynman — save that it contains a Vera CPU. Nvidia plans to deliver Feynman, which is able to succeed Rubin Vera, to market someday in 2028.
Up to date 3/18 3:07 p.m. Pacific: An earlier model of this story indicated that Vera Rubin had “tens of terabytes” of reminiscence. Actually, it has “tens of gigabytes” of reminiscence. We remorse the error.
About The Author
