By admin 
On Tuesday at Nvidia’s GTC 2025 conference in San Jose, California, CEO Jensen Huang revealed plenty of new AI-accelerating GPUs the company plans to launch over the approaching months and years. He moreover revealed further specs about beforehand launched chips.
The centerpiece announcement was Vera Rubin, first teased at Computex 2024 and now scheduled for launch throughout the second half of 2026. This GPU, named after a well-known astronomerwill perform tens of terabytes of memory and comes with a custom-made Nvidia-designed CPU known as Vera.
In response to Nvidia, Vera Rubin will ship very important effectivity enhancements over its predecessor, Grace Blackwell, notably for AI teaching and inference.
Specs for Vera Rubin, launched by Jensen Huang all through his GTC 2025 keynote.
Vera Rubin choices two GPUs collectively on one die that ship 50 petaflops of FP4 inference effectivity per chip. When configured in a full NVL144 rack, the system delivers 3.6 exaflops of FP4 inference compute—3.3 events better than Blackwell Extraordinarily’s 1.1 exaflops in the identical rack configuration.
The Vera CPU choices 88 custom-made ARM cores with 176 threads linked to Rubin GPUs by means of a high-speed 1.8 TB/s NVLink interface.
Huang moreover launched Rubin Extraordinarily, which might adjust to throughout the second half of 2027. Rubin Extraordinarily will use the NVL576 rack configuration and have specific 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 diploma, Rubin Extraordinarily will current 15 exaflops of FP4 inference compute and 5 exaflops of FP8 teaching effectivity—about 4 events further extremely efficient than the Rubin NVL144 configuration. Each Rubin Extraordinarily GPU will embrace 1TB of HBM4e memory, with the entire rack containing 365TB of fast memory.