Nvidia Jensen Huang CEO arrives to attend the opening ceremony of Siliconware Precision Industries Co. (Spil), Tan Ke Plants in Taichung, Taiwan on January 16, 2025.
Ann Wang | Reuters
Nvidia He has announced a new chips for the construction and deployment of artificial intelligence models on his annual GTC conference on Tuesday.
Executive director Jensen Huang was discovered by Blackwell Ultra, a chips family in the second half of this year, Like Vera RubinThe next -generation or GPU graphic processing unit, which is expected to be delivered in 2026.
The sale of NVIDIA has increased more than six times as its business has been transformed by the publication of Openai’s Chatgpt at the end of 2022. This is because his “big GPUs” are most of the market for the development of advanced AI, a process called training.
Software programmers and investors are watching new company chips carefully to see if they offer enough additional performance and efficiency to convince the largest end customers of the company – Cloud Company, including Microsoft,, Google and Amazon – To continue spending billions of dollars to build data centers based on Nvidia chips.
“This last year is in which almost the whole world has become involved. A computer request, the AI Scalar Law, is more resistant and things are more hyper-stroke,” Huang said.
The announcements on Tuesday are also a test of the new annual Cadence NVIDIA. The company strives to publish new Chip families every year. Before AI BUM, Nvidia has published new Chip architecture every other year.
The GTC conference San Jose, California, is also a display of strength for Nvidia.
The second event, the second NVIDIA conference from Pandemia, is expected to have 25,000 students and hundreds of companies discussing the ways they use the AI hardware. This includes Waymo, Microsoft and Fordamong others. General Motors also announced that he would use Nvidia’s service for his of the next generation vehicles.
Nvidia will also show her other products and services at the event.
For example, Nvidia has announced new laptops and desktops using her chips, including two computers ai that will be able to start large AI models such as Llam or Deepseek. The company also announced updates of its network parts to connect hundreds or thousands of GPU, so they work as one, as well as a software package called Dynamo that helps users to maximize their chips.
Jensen Huang, co -founder and director of Nvidia Corp., speaks during the Nvidia GPU Technology Conference (GTC) in San Joseu, California, USA, Tuesday, March 18, 2025.
David Paul Morris | Bloomberg | Getty Images
Vera Rubin
Nvidia expects that in the second half of 2026. Delivery systems will begin in their next generation GPU family.
The system has two main components: CPU, called Vera and a new GPU design, called Rubin. The name is named after Astronomer Vera Rubin.
Vera Nvidia is the first custom CPU design, the company said, based on the basic design that they named Olympus.
Earlier when needed by CPUs, Nvidia used the design outside the shelf from Hand. Companies that have developed adapted to ARM designs, such as Qualcomm -Ai Apple, say they can be more adapted and unlock better performance.
The adapted design of Vera will be twice as fast as the CPU used in last year’s Grace Blackwell Chips, the company said.
When paired with Vera, Rubin can manage 50 petaflops while for the conclusion, more than double 20 petaflops for current Blackwell chippoves. Rubin can also support as many as 288 gigabytes of fast memory, which is one of the main specifications that AI developers watch.
Nvidia also changes what he calls GPU. Rubin is actually two GPUs, Nvidia said.
The Blackwell GPU, which is currently on the market, is actually two separate chips that are assembled together and made to work as one chip.
Starting with Rubin, Nvidia will say that when he combines two or more matrices to make one chip, he will call them a separate GPU. In the second half of 2027, Nvidia plans to issue a “Rubin Next” chip that combines four matrices to make one chip, double the speed of rubies, which will be stated as four GPUs.
Nvidia said she would come to a stand called Vera Rubin NVL144. The previous versions of Nvidia’s rack were called NVL72.
Chip architecture after Rubin will be appointed by physicist Richard Feynman, Nvidia said on Tuesday, continuing his tradition of appointing a chip to scientists. Feynman chips of Nvidia are expected to be available in 2028, according to a diamond presented by Huang.
Jensen Huang, co -founder and director of Nvidia Corp., speaks during the Nvidia GPU Technology Conference (GTC) in San Joseu, California, USA, Tuesday, March 18, 2025.
David Paul Morris | Bloomberg | Getty Images
Blackwell Ultra
Nvidia also announced new versions of her Blackwell Family of Chips called Blackwell Ultra.
This chip will be able to produce multiple token per second, which means that the chip can generate more content at the same time as its predecessor, the company in Brifing announced.
Nvidia says this means that cloud service providers can use Blackwell Ultra to offer superior AI services for time sensitive applications, allowing them to earn as much as 50 times more than new chips like Hopper, which was delivered in 2023.
Blackwell Ultra will come in two -paired version with Nvidia Arm CPU, called GB300, and a version with only a GPU, called B300. It will also be found in versions with eight GPUs with one knife server and a 72 black -based standing version.
Four four of the cloud companies have arranged three times the number of Blackwell chips as a tank chips, Nvidia said.
Deepseek
The Chinese model Deepseek R1 may have scared Nvidia investors when it was published in January, but Nvidia accepted the software. A chip manufacturer will use a model to compare several of its new products.
Many AI observers said that Deepseek’s model, which allegedly demanded fewer chips than models made in the United States, threatened Nvidia’s business.
But Huang said earlier this year that Deepseek was actually a good sign for Nvidia. This is because Deepseek uses a procedure called “reasoning”, which requires more computer power to provide users with better answers.
The new Blackwell Ultra chips are better for reasoning models, Nvidia said.
He has developed his chips to make more effective conclusions, so when new reasoning models require more computer power at the time of implementation, Nvidia’s chips will be able to handle it.
“In the last 2 to 3 years, a major breakthrough has occurred, the fundamental progress of artificial intelligence has occurred. We call it an agent AI,” Huang said. “This can reason to answer or how to solve the problem.”
WATCH: Nvidia begins its GTC Conference: Committee discusses how to trade it
