On Tuesday at Nvidia’s GTC 2025 convention in San Jose, California, CEO Jensen Huang revealed a number of new AI-accelerating GPUs the corporate plans to launch over the approaching months and years. He additionally revealed extra specs about beforehand introduced chips.
The centerpiece announcement was Vera Rubin, first teased at Computex 2024 and now scheduled for launch within the second half of 2026. This GPU, named after a famous astronomer, will characteristic 288 gigabytes of reminiscence and comes with a customized Nvidia-designed CPU referred to as Vera.
In accordance with Nvidia, Vera Rubin will ship vital efficiency enhancements over its predecessor, Grace Blackwell, notably for AI coaching and inference.
Specs for Vera Rubin, introduced by Jensen Huang throughout his GTC 2025 keynote.
Vera Rubin options two GPUs collectively on one die that ship 50 petaflops of FP4 inference efficiency per chip. When configured in a full NVL144 rack, the system delivers 3.6 exaflops of FP4 inference compute—3.3 occasions greater than Blackwell Extremely’s 1.1 exaflops in an analogous rack configuration.
The Vera CPU options 88 customized ARM cores with 176 threads linked to Rubin GPUs by way of a high-speed 1.8 TB/s NVLink interface.
Huang additionally introduced Rubin Extremely, which is able to comply with within the second half of 2027. Rubin Extremely will use the NVL576 rack configuration and have particular person GPUs with 4 reticle-sized dies, delivering 100 petaflops of FP4 precision (a 4-bit floating-point format used for representing and processing numbers inside AI fashions) per chip.
On the rack stage, Rubin Extremely will present 15 exaflops of FP4 inference compute and 5 exaflops of FP8 coaching efficiency—about 4 occasions extra highly effective than the Rubin NVL144 configuration. Every Rubin Extremely GPU will embody 1TB of HBM4e reminiscence, with the entire rack containing 365TB of quick reminiscence.
