Nafidia announces the flakes of Blackwell Ultra and Vera Robin AI

The CEO of Nvidia Jensen Huang arrives to attend the opening ceremony of Precision Industries Co. (Spil) in Taichung, Taiwan, January 16, 2025.

Ann Wang Reuters

Nafidia New chips for building and publishing artificial intelligence models announced at the annual GTC conference on Tuesday.

CEO Jensen Huang Blackwell Ultra, a family of shipping chips in the second half of this year, as well as As Vera RobinThe graphics processing unit from the next generation of the company, or the graphics processing unit, is expected to be charged in 2026.

NVIDIA sales have risen more than six times since its business has turned through the Openai’s Chatgpt in late 2022. That is because “GPU Big Big” has most of the market to develop AI advanced, a process called training.

Software and investors are closely seen by the company’s new chips to see if they are providing an additional and efficient performance to persuade the company’s largest customer – cloud companies including Microsoftand Google and Amazon To continue spending billions of dollars to build NVIDIA databases.

“This last year is the place where the whole world has been involved. The mathematical condition is the law of scaling in artificial intelligence, more flexible, and in fact, it is accelerated,” Huang said.

Tuesday’s ads are also a test of the new annual version of NVIDIA. The company seeks to announce new chip families on the basis of every year. Before the Amnesty International Break, NVIDIA released a new chips structure every year.

The GTC Conference in San Jose, California is a force for NVIDIA.

It is expected that the event, which is the second personal NVIDIA conference since the epidemic, is 25,000 of the attendees and hundreds of companies that discuss the methods in which they use the company’s devices for artificial intelligence. It includes Waymo and Microsoft and FordAmong other things. General Motors She also announced that she will use her NVIDIA service Vehicles of the next generation.

Nafidia said on Tuesday, the structure of the chip after Robin, which will be called the name of the physicist Richard Fainman. Feynman foil is expected to be available from NVIDIA in 2028, according to a SUANG chip.

NVIDIA will also display its products and other services in this event.

For example, NVIDIA announced laptops and new desktop computers using their chips, including computers that focus on artificial intelligence called called DGX Spark and DGX station This will be able to run large models of artificial intelligence like Llama or Deepseek. The company has also announced updates for its network spare parts to connect hundreds or thousands of graphics processing together so that they act as a DYNAMO software package that helps users get the utmost benefit from their chips.

Jensen Huang, co -founder and CEO of Nvidia Corp, speaks during the Nvidia Gpu Technology (GTC) conference in San Jose, California, United States, on Tuesday, March 18, 2025.

David Paul Morris Bloomberg Gety pictures

Vera Robin

NVIDIA expects to start charging systems on the GPU from the next generation in the second half of 2026.

The system has two main components: the CPU, called Vera, and a new GPU design, called Rubin. Bassem named Astronomy Vera Robin.

The company said Vera is the first designated design in the CPI in NVIDIA, and is based on a basic design that they named Olympus.

Previously when CPU needs, NVIDIA used a design out of the shelf arm. Custom Arm Core designs, such as Qualcomm and Apple, says they can be more customized and open better performance.

The company said that the design of VIRA will be twice the speed of the CPC in Grace Blackwell Chips last year.

When associated with Vera, Rubin can manage 50 PETAFLOPS while inference, more than twice 20 PETAFLOPs for the current Blackman Black Cools. Robin can also support up to 288 GB of fast memory, which is one of the basic specifications that artificial intelligence developers see.

NVIDIA also makes a change to what you call the graphics processing unit. Nafidia said that Robin is actually graphics processing units.

The GPU Blackwell, currently on the market, is two separate chips that were assembled together and made to work as one bit.

From Robin, Nvidia will say that when a two or more deaths combine to create one chip, they will indicate it as separate official processing units. In the second half of 2027, NVIDIA plans to launch the “Rubin Next” chip that combines four deaths to make one slide, the Robin speed doubles, and it will indicate this as four graphics processing units.

Nafidia said this will come in a shelf called Vera Rubin Nvl144. It was called previous versions of the NVIDIA NVL72 shelf.

Jensen Huang, co -founder and CEO of Nvidia Corp, speaks during the Nvidia Gpu Technology (GTC) conference in San Jose, California, United States, on Tuesday, March 18, 2025.

David Paul Morris Bloomberg Gety pictures

Blackwell Ultra

NVIDIA has also announced new versions of the Blackwele family of chips called Blackwell Ultra.

The company said in a press conference that this slide will be able to produce more symbols per second, which means that the slide can generate more content in the same time as it preceded.

This means that the cloud service providers can use Blackwell Ultra to provide a distinctive AI service for time -sensitive applications, allowing them to achieve up to 50 times the revenues of new chips such as the Hopper generation, which was shipped in 2023.

Blackwell Ultra will come in a version with two associated with the ARM NVIDIA CPU, called GB300, and a version with the graphics processing unit only, called B300. It will also come in versions with eight graphics processing units in one server blade and version of a shelf with 72 Blacks chips.

Nafidia said that the four best cloud companies have deployed three times the number of Black Cell chips as Hopper’s Frames.

Dibsic

The Chinese Deepseek R1 model may have been afraid of NVIDIA investors when it was released in January, but NVIDIA adopted the program. The chips maker will use the model to measure many of its new products.

Several artificial intelligence monitors said that the Deepseek model, which requires reports of fewer chips of models made in the United States, threatened NVIDIA’s works.

But Huang said earlier this year that Dibsic was actually a good sign of Nafidia. This is because Deepseek uses a process called “thinking”, which requires more computing power to provide users with better answers.

Nafidia said the new Blackwell Ultra chips are better for thinking models.

Its chips have developed to conclude them more efficiently, so when new thinking models require more computing power at the time of publication, NVIDIA chips will be able to deal with them.

“In the past two years to 3 years, there has been a major breakthrough, a basic progress in artificial intelligence. We call it artificial intelligence agent,” Huang said. “It can be the reason how to answer or how to solve a problem.”