On Tuesday at Nvidia’s GTC 2025 conference in San Jose, California, CEO Jensen Huang exposed a number of brand-new AI-accelerating GPUs the business prepares to launch over the coming months and years. He likewise exposed more requirements about formerly revealed chips.

The focal point statement was Vera Rubin, very first teased at Computex 2024 and now set up for release in the 2nd half of 2026. This GPU, called after a well-known astronomer, will include 288 gigabytes of memory and features a customized Nvidia-designed CPU called Vera.

According to Nvidia, Vera Rubin will provide considerable efficiency enhancements over its predecessor, Grace Blackwell, especially for AI training and reasoning.

Vera Rubin includes 2 GPUs together on one die that provide 50 petaflops of FP4 reasoning efficiency per chip. When set up in a complete NVL144 rack, the system provides 3.6 exaflops of FP4 reasoning calculate– 3.3 times more than Blackwell Ultra’s 1.1 exaflops in a comparable rack setup.

The Vera CPU includes 88 custom-made ARM cores with 176 threads linked to Rubin GPUs through a high-speed 1.8 TB/s NVLink user interface.

Huang likewise revealed Rubin Ultra, which will follow in the 2nd half of 2027. Rubin Ultra will utilize the NVL576 rack setup and function specific GPUs with 4 reticle-sized passes away, providing 100 petaflops of FP4 accuracy (a 4-bit floating-point format utilized for representing and processing numbers within AI designs) per chip.

At the rack level, Rubin Ultra will offer 15 exaflops of FP4 reasoning calculate and 5 exaflops of FP8 training efficiency– about 4 times more effective than the Rubin NVL144 setup. Each Rubin Ultra GPU will consist of 1TB of HBM4e memory, with the total rack consisting of 365TB of quick memory.