① Huang Renxun introduced the new product 'Blackwell Ultra' during his keynote speech at GTC, and previewed the next-generation chip 'Rubin'; ② Blackwell Ultra is based on the Blackwell architecture, integrating NVIDIA's GB300 NVL72 rack-level solution and NVIDIA's HGX B300 NVL16 system.
Financial Network reported on March 19 (Editor Zhao Hao) that on Tuesday local time (March 18), $NVIDIA (NVDA.US)$ CEO Jensen Huang launched the new product 'Blackwell Ultra' during the GTC keynote speech and previewed the company's next-generation chip 'Rubin'.
Huang Renxun stated that almost the entire world participated in the construction of AI Datacenters last year, stating that "computing demand - namely the AI scaling law - is more elastic and is rapidly increasing in speed."
"AI has made tremendous leaps - inferencing and agent-based AI require orders of magnitude higher computing power. We designed Blackwell Ultra for this moment, as a powerful single platform capable of effortlessly and efficiently performing pre-training, post-training, and inferencing tasks for AI."
According to NVIDIA's official website, Blackwell Ultra is based on the Blackwell architecture launched by the company a year ago, combining the NVIDIA GB300 NVL72 rack-level solution and NVIDIA HGX B300 NVL16 systems.
The GB300 NVL72 integrates 72 Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs based on Arm Neoverse architecture, forming a massive single GPU designed for scaled inferencing during testing.
Through the GB300 NVL72, AI models can leverage the platform's enhanced computing power to explore different solutions and break down complex requests into multiple steps, leading to higher-quality responses.
The press release states that the HGX B300 NVL16 is 11 times faster than Hopper in terms of large language model (LLM) inferencing, offering 7 times the computing power and 4 times the memory, enabling breakthrough performance improvements for the most complex workloads such as AI inferencing.
NVIDIA stated that the performance of the new rack-level solution is 1.5 times that of the NVIDIA GB200 NVL72; compared to factories built using NVIDIA Hopper, Blackwell's AI factory revenue opportunity has increased by 50 times.
Jensen Huang stated in his speech that "the performance of Blackwell in inference models is 40 times that of Hopper." He also mentioned that the company is fully focused on the production of Blackwell, transitioning to Blackwell Ultra in the second half of the year.
The press release mentioned,$Cisco (CSCO.US)$、 $Dell Technologies (DELL.US)$ 、$HP Inc (HPQ.US)$、 $LENOVO GROUP (00992.HK)$ 、$Super Micro Computer (SMCI.US)$Various servers based on the Blackwell Ultra products will be provided.$Amazon (AMZN.US)$AWS, Google Cloud,$Microsoft (MSFT.US)$Azure and others will also offer related Cloud Computing Services for Blackwell Ultra.
It is worth mentioning that the Blackwell Ultra chips can also be purchased individually. Jensen Huang launched a desktop computer called "DGX Station" at the event, which features a single GB300 Grace Blackwell Ultra and 784GB of memory, and ConnectX-8 SuperNIC that supports up to 800Gb/s networking.
At the event, Jensen Huang previewed the next generation AI chip architecture "Rubin" and the custom CPU "Vera," which are positioned against the current Blackwell and Grace CPUs. Vera Rubin is scheduled for release in the second half of 2026, and Rubin Ultra in the second half of 2027.
Vera Rubin is a renowned astronomer in the USA, who made groundbreaking advances in dark matter research, fundamentally changing human understanding of the universe.
Vera integrates 88 custom Arm cores, 176 threads, and 1.8TBp/s NVLink-C2C. Rubin contains two GPUs and achieves 50PF in FP4 precision inference performance, while also supporting up to 288GB of fast memory—core specifications of interest to AI developers.
The Vera Rubin NVL144 system can perform 3.6 EF of FP4 inference and 1.2 EF of FP8 training, with a performance approximately 3.3 times that of the GB300 NVL72.
The Rubin Ultra NVL576 system, launched in 2027, can perform 15 EF of FP4 inference and 5 EF of FP8 training, with a performance approximately 14 times that of the GB300 NVL72.
Jensen Huang also announced that the next generation after Rubin will be named Feynman, after the famous American physicist Richard Feynman. According to NVIDIA's roadmap, the Feynman architecture will debut in 2028.
As of the time of writing, $NVIDIA (NVDA.US)$ Closed down over 3%, at $115.43.
Editor/Rocky
