Nvidia announced Blackwell Ultra and Vera Rubin AI Chips

Nvidia In its annual GTC conference on Tuesday announced new chips for the construction and deployment of Artificial Intelligence Models.

CEO Jensen Huang, in the second half of this year, revealed the family of chips shipping Blackwell Ultra, as well as the company’s next generation graphics processing unit, or GPU, Vera Rubin, a ship expected to ship in 2026.

The sale of NVidia is more than six times as its business was changed by the release of Openai’s chatgpt at the end of 2022. This is because most of its “Big GPU” has a market to develop advanced AI in the market, a process called training.

Software developers and investors are closely looking at the company’s new chips whether they provide enough additional performance and skills to explain the company’s biggest end customers – including cloud companies Microsoft, Google And Heroic – Continue spending billions of dollars to create data centers based on Nvidia chips.

“This is the place last year where almost the whole world joined. Computational requirement, AI’s scaling law, is more flexible, and in fact, hyper-academic,” said Huang.

Tuesday’s announcements are also tested for Nvidia’s new annual release pool. The company is trying to announce new chip families based on every year. Prior to the AI ​​Boom, Nvidia released new chip architecture every other year.

The GTC conference in San Jose, California is also a performance of strength for NVidia.

The event is expected to be present at Nvidia’s second-opinion conference since the epidemic, and hundreds of companies discuss ways to use the company’s hardware for AI. Which includes Waymo, Microsoft and PaidAmong others. General Motors It also announced that it would use NVIDIA service for its next generation of vehicles.

After Rubin, the chip architecture will be named after the physicist Richard Fenman, Navidia said on Tuesday, “After scientists continued their tradition of naming chip families.” According to a slide displayed by Huang, NVIDIA’s fenman chips are expected to be available in 2028.

The NVIDIA will also demonstrate its other products and services in the event.

For example, Nvidia announced new laptops and desktop using its chips, including two AI-focused PCs, called DGX Spark and DGX stations that will be able to run large AI models such as Lama or Deepsek. The company announced updates in its networking parts to tie hundreds or thousands of GPUs together, so that they work as one, as well as a software package called dynamo that helps users to get the most from their chips.

NVIDIA is expected to launch a shipping system on its next generation GPU family in the second half of 2026.

The system has two main components: a CPU, called Vera, and a new GPU design, called Rubin. It is named after astronomer Vera Rubin.

Vera is Nvidia’s first custom CPU design, the company said, and it is based on a core design she has named Olympus.

Earlier when it was needed CPU, Nvidia used an off-the-shelf design HandCompanies that have developed custom arm core designs, such as Qualcomm and Apple, say they can be more sewn and unlock better performance.

The company said the custom vera design would double as a CPU used in last year’s Grace Blackwell Chips.

When combined with vera, the rubin can manage 50 petaphylops, doubling the company’s current blackwell chips more than 20 petaphylops. Rubin can also support Fast memory as 288 gigabytes, one of the core glasses watching AI developers.

NVIDIA is also making a change on what a GPU is called. Rubin is actually two GPU, Nvidia said.

Blackwell GPU, which is currently in the market, is actually two different chips that were collected together and made to work as a chip.

Starting with Rubin, Nvidia would say that when it dies two or more to make the same chip, it will refer to them as separate GPUs. In the second half of 2027, Nvidia planned to release a “Rubin Next” chip, which dies four to create a single chip, which doubles the speed of the rubin, and it will refer to as four GPUs.

NVIDIA said that Vera Rubin would come in a rack called NVL144. The previous versions of the NVIDIA rack were called NVL72.

NVIDIA also announced new versions of its Blackwell family chips, which she calls Blackwell Ultra.

This chip will be able to produce more tokens per second, meaning that the chip can produce more materials at the same time as its predecessor, the company said in a briefing.

Nvidia says that this means that cloud provider Blackwell Ultra can be used to offer premium AI service for time-sensitive applications, allowing them to create 50 times more revenue than new chips as a hopper generation, which was sent to 2023.

The Blackwell Ultra will come in a version with two an NVidia Arm CPU, called the GB300, and just a version with GPU, called the B300. It will also come in versions with a rack version with eight GPUs and 72 blackwell chips in single server blade.

NVIDIA said that the top four cloud companies have deployed the number of blackwell chips three times as hopper chips.

On release in January, China’s Dipsek R1 model may have scared NVidia investors, but NVidia has adopted the software. Chipmaker will use models to benchmark many of its new products.

Several AI observers stated that the models of the Deepsek, who were allegedly required less chips than the model made in the US, threatened NVidia’s business.

But Huang said earlier this year that Deepsek was actually a good sign for Nvidia. This is because the lamp uses a process called “Reasoning”, which requires more computing power to provide better answers to users.

New Blackwell Ultra Chips is better for the Reasoning Model, Nvidia said.

This has developed its chips to develop more efficiently, so when the new logic model requires more computing power at the time of deployment, NVidia’s chips will be able to handle it.

“In the last 2 to 3 years, there has been a major success, a fundamental advance in artificial intelligence. We call it agent AI,” said Huang. “This may be the reason about how to do or how to do a problem.”

Watch: NVIDIA discontinues its GTC conference: the committee argued how to do business