NVIDIA announces next-generation AI-specialized GPU architectures 'Blackwell Ultra', 'Vera Rubin' and 'Rubin Ultra', Rubin is 900 times more powerful than Hopper and costs only 3%

NVIDIA has announced details of its next-generation architectures for AI data centers: Blackwell Ultra , Vera Rubin , and Rubin Ultra . Products equipped with Blackwell Ultra are scheduled to begin shipping in the second half of 2025 and will have 1.5 times the performance of the previous generation of products.
NVIDIA Blackwell Ultra AI Factory Platform Paves Way for Age of AI Reasoning | NVIDIA Newsroom
GTC March 2025 Keynote with NVIDIA CEO Jensen Huang - YouTube
NVIDIA has announced the GB300 NVL72 , an AI processing system for data centers that uses Blackwell Ultra. The GB300 NVL72 is equipped with 72 Blackwell Ultra GPUs and 36 Arm-based CPUs called NVIDIA Grace .

In addition, Blackwell Ultra is equipped with 40TB of high-speed memory and is seamlessly integrated with high-speed communication systems such as

When using the server board '

Also, the Llama 3.1 405B will learn four times faster.

NVIDIA also released information about 'Vera Rubin' and 'Rubin Ultra'. In the second half of 2026, the Vera Rubin NVL144, a data center system equipped with a Vera Rubin GPU, is scheduled to be released. The Vera Rubin NVL144 is equipped with 75TB of high-speed memory, providing 3.3 times the performance of the GB300 NVL72.
In addition, the Rubin Ultra NVL576, equipped with a Rubin Ultra GPU, will be released in the second half of 2027. The Rubin Ultra NVL576 will have a memory capacity of 365TB and 14 times the processing power of the GB300 NVL72.

Here's an image of the GB300 NVL72 GPU die: It has 130 trillion transistors.

Below is an image of the Rubin Ultra NVL576 GPU die. The size of the GPU is huge, and the number of transistors has reached 1,300 trillion.

Below is a comparison of the performance and cost of the Hopper, Blackwell, and Rubin generations. Blackwell has 68 times the performance of Hopper, and the cost is 13%. Rubin has 900 times the performance of Hopper, and the cost is 3%.

Related Posts:
in Hardware, Posted by log1o_hf