On the stage at the NVIDIA GTC 2025 conference in San Jose On Tuesday, CEO JenSen Huang announced a large number of new graphics processing units that are in the company’s product pipeline during the next few months.
Perhaps the most important is Vera Robin. The Vera Rubin, released in the second half of 2026, will include dozens of memory and the designer CPI designed dedicated called Vera. Vera Rubin performs a great performance compared to its predecessor, Grace Blackwell, NVIDIA claims, especially in the tasks of inference and artificial intelligence training.
When pairing with Vera, Rubin – which is the processing of graphics processing units in one, technically – can manage up to 50 PETAFLOPs while in reasoning (i.e., running artificial intelligence models), more than twice 20 PETAFLOPs for the current Blackwele slices in NVIDIA. Moreover, Vera is about twice the speed of the CPI in Nvidia’s Grace Blackwell GPU.
Robin Robin will follow in the second half of 2027, a group of four graphics processing units in one package that provides up to 100 Petaflops of performance.
On the near horizon – H2 2025 – NVIDIA will launch Blackwell Ultra, a graphics processing unit that will come in many configurations. One sophisticated segment will provide the same 20 PETAFLOPs of artificial intelligence like Blackweell, but with 288 GB of memory – a height of 192 GB in vanilla Blackwell.
On the distant horizon is Feynman graphics processing units. During the keynote, Huang gave some details about Feynman architecture, which was named after the American theoretical physicist Richard Fainman – except that it features the Ferira CPU. NVIDIA plans to bring Feynman, which will succeed Rubin Vera, to the market at some point in 2028.
Updated 3/18 3:07 PM Pacific: A previous version of this story indicated that Vera Rubin had “dozens of TERABYTES” of memory. In fact, it has “dozens of GB” of memory. We regret the error.
https://techcrunch.com/wp-content/uploads/2025/03/GettyImages-2205211212-1.jpg?resize=1200,800
Source link