Forward-loving: NVIDIA CEO Jensen Huang unveiled a strong lineup of the AII-e-e-recovery GPU at the company’s 2025 GPU technology conference, including Blackwell Ultra B300, Vera Rubin and Rubin Ultra. These GPU AI are designed to increase performance, especially in estimates and training functions.
Blackwell Ultra B300, set for release in the second half of 2025, increases the memory capacity from 192GB to HBM3E to 288GB and promotes 50% in the calculation of dense FP4 tensor compared to Blackwell GB200.
These enrichrs support large AI models and improve estimates for framework such as Deepsek R1. In a full NVL72 rack configuration, Blackwell Ultra will distribute 1.1 exaflops of dense FP4 Infererance Compute, marking a significant jump on the current Blackwell B200 setup.
Blackwell Ultra B300 is not just a standalone GPU. With core B300 unit, NVIDIA is launching the new B300 NVL16 server rack solutions, GB300 DGX station and GB300 NV72L Full Rack System.
The combination of eight NV72L racks is characterized by a full blackwell ultra DGX superpod (depicted above), 288 grace CPU, 576 Blackwell Ultra GPU, 300TB of HBM3E Memory, and an impressive 11.5 exaflops of FP4 compute power. These systems can be added to each other to create a large -scale supercomputer, called Nvidia “AI factories”.
Initially Computex was teased on 2024, the next-gene vera Rubin GPUS is expected to be launched in the second half of 2026, especially the AI training and estimates have improved adequate performance.
Vera Rubin has tensual terabytes memory and is paired with a custom Nvidia-designed CPU, Vera, including 88 custom arm core with 176 threads.
GPU integrates two chips when a single dies, receiving 50 petflops of estimates per chip per chip. In a full NVL144 rack setup, Vera Rubin FP4 can distribute 3.6 exaflops of Infererance Compute.
On the architecture of Vera Rubin, the building, Rubin Ultra has been slated for release in the second half of 2027. This will use the NVL576 rack configuration, dying four ratical-shaped with each GPU, providing 100 petflops per chip FP4 accuracy.
Rubin Ultra promised 5 exaflops of 15 exaflops and FP8 training performance of FP4 Infererance Compute, crossing the capabilities of Vera Rubin. Each Rubin Ultra GPU will include 1TB of HBM4E memory, which contributes to 365TB fast memory throughout the rack.
Nvidia had expected the next generation GPU architecture to debut in 2028 with Wera CPU in 2028. While the details are rare, Fenman is anticipated to pursue Nvidia’s AI computing capabilities.
During his keynote speaker, Huang underlined NVidia’s ambitious vision for AI, describing data centers as “AI factories” that produce tokens processed by the AI model. He also highlighted the ability to power “physical AI”, taking advantage of NVidia’s software platforms to train the AI model in a virtual environment for real -world applications.
In the future of roadmap computing of Nvidia, these GPUs are important, which emphasizes the need for computational power to keep pace with AI progress. This strategy comes when NVIDIA aims to assure investors after the recent market, to create the success of its blackwell chips.
