At its GTC Year Developer summit in San Jose, Calif, Calif. Today (Marif 18), Nvida (NVDA) is revealed to two future GPU buildings: blackwell ultra and Rubin. During his opening keynote, the CEO JENDER Huang set up the promotion of Chipmaker prominence to move AI into a industrial period. Huang called GTC “Chief Acher of Ai” One keyza: “In this large vessel, everyone is successful,” he said.
Blackwell Ultra is expected to launch in the second half of 2025. The following, Rubin, Sheliz by late 2026, followed by Rubin Ultra in the highest level of 2027.
Knowing Ablackwell Ultra and Rubin
Currently in production, a series of advanced Blackkwell of the Blackwell Chips are revealed in the GTC annual year, 40 40 times highly powerful Hopsper Chips, said Huang.
Blackwell Ultra will put eight memory stacks for 12-HI HBM4e, provides the 288GB of onboard memory. Construction will include the NVLIK 72, upgraded Top Technical Technology designed to facilitate communication between GPUS and CPUS-most important consideration of a large AI and acquisition.
“NVLink links many GPUs, which turn it into one GPU,” explains Huang. “It deals with the measuring problem by enabling the main compatible computing.”
In addition, the company reveals the NVIA RTX Pro 6000 Blackwell Server Edition, engineering engineering and multimodial activity for multilateral and scientific content and science computer. With the 96GB of GDDR7 memory and the Multi-Instance GPU Technology support, RTX Pro 6000 is designed for ADDAND ADVAND AI.
Blackwell Follower of Blackwill, Rubin, is named at Astronoms Vera Rubin, who found the existence of a black substance in space. The first type of rubin chip is expected to reach 50 petAflops during the AI ​​model murder. The strongest powerful Rubin can move until 100 PetaFLOPS WORKING, Huang called “a major step forward” to work and work.
The first arrival of black chips and racks reported that there were experiences of excessive heatleading certain customers to reduce their order. The recently introduced-dropped-dropped program in the Blackwell 200 NVL72 program deals with the worries, providing 30 times for the actual trillion-language models. It can build up to 12,000 tokens for a second – the basic units are processed AI models – Awareness of the amazing speed of both training and monitoring.
“If you want your AI to be skillful, it should produce many tokens. That requires a major bandwidth, floating works and cow,” said Huang. He also explained that AI models were like a Deepseeeke’s R1 model who needed tokens 20 times and the highest 105 powers.
Invidia also works closely with Taiwan TSMC to promote data packing technology, movement that can be highly increasing GPU generations.
Nvidia’s street passes Rubin
Huang also described the NVInia road by Rubin. “The next generation of buildings after Rubin will be called Feynman,” said Feynman HBM is already developing and separated by Richard Feynman, this is expected to oppose the AI.
The announcements occur in the best nvidia and the best nvidia are expected resultsdriven by a passing demand for its GPU. Despite the increasing competition from rival and geopolitical uncertainty includes international export restrictions on the semiconductor, according to an 80% market report, according to report In Nasdaq.
