Nvidia announces Blackwell Ultra and Vera Rubin AI chips

Nvidia Jensen Huang CEO comes to attend the precision Industries Co. Silicon Opening Ceremony. (Spil) “San Ces Play” in Tychchung, Taiwan, January 16, 2025.

Anne van | Reuters

Nvidia At their annual GTC conference on Tuesday, they announced new chips to create and deploy artificial intelligence models.

CEO Jensen Juan showed that Blackwell Ultra, the Chips family, in the second half of this year, Like Faith RubinThe next -generation graphic unit, or GPU, is expected to be sent in 2026.

Nvidia sales grew more than six times as its business was converted to Chatgpt Openai in late 2022. This is because its “big graphic processors” have most of the market for the development of Advanced AI, a process called Training.

Software developers and investors closely monitor the new Microsoft. Google and Amazon – Continue to spend billions of dollars on the construction centers based on Nvidia chips.

“Last year, when the whole world is almost involved. Compuration, the law on scaling II, more elastic, and in fact, hyper-scratching,” Juan said.

Tuesday ads are also a test of Nvidia’s new annual cadence. The company seeks to announce new chips daily. Before the Boom of AI Nvidia, new architecture chips every second year.

The GTC Conference in San Khase, California, is also an indicator of Nvidia.

It is expected that at the second NVIDIA conference of the pandemic will be expected to be 25,000 visitors and hundreds of companies discussing the company’s equipment for AI equipment. These include Waymo, Microsoft and FreodAmong others. General Motors also announced that would use the NVIDIA service for its The next -generation vehicles.

The architecture of the chips after Rubin will be named after the physicist Richard Feinman, Nwidia said on Tuesday, continuing the tradition to call families chips after scientists. Feynman Nvidia chips are expected to be available in 2028, Slide reports.

Nvidia will also demonstrate its other products and services at the event.

For example, Nvidia has announced new laptops and desktops, using their chips, including two PCs oriented by AI, called Spark and DGX Station, which will be able to launch large AI models such as Llama or Deepseek. The company also announced updating its network parts to bind hundreds or thousands of graphic processors together to work as one as well as a software pack called Dynamo, which helps users to use their chip as much as possible.

Jensen Huang, co-founder and chief executive director of Nvidia Corp., performs during the Nvidia GPU Technology (GTC) conference in San Jose, California, USA, on Tuesday, March 18, 2025.

David Paul Morris | Bloomberg | Gets the image

Faith Rubin

Nvidia expects to start delivery systems in its next -generation family in the second half of 2026.

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

The company – Vera is the first design of the Nvidia processor, and it is based on the main design they called Olympus.

Used to be when it needed Arm. Companies that have developed non -standard ARM designs such as Qualcomm and Apple say they can be more developed and unlock better performance.

Vera Design will be twice as faster than the processor used in last year’s Grace Blackwell chips, the company reports.

In paired with faith, Rubin can manage 50 Petaflops, drawing conclusions, more than 20 petoflops for current Blackwell companies. Rubin can also support as much as 288 rapid memory gigabytes, which is one of the main specifications observed by AI developers.

Nvidia also makes changes to what he calls a graphic processor. Rubin is actually two graphic processors, Nvidia said.

GPU Blackwell, which is currently on the market, actually has two separate chips that have been assembled together and forced to work as a single chip.

Starting with Rubin, Nvidia will say that when it combines two or more stamps to make one chip, it will call them separate graphic processors. In the second half of 2027, Nvidia plans to release the Rubin Next, which combines four stamps to make one chip, double the Rubin speed, and it will call it four graphic processors.

Nvidia said it would come to a counter called Vera Rubin NVL144. Previous versions of the Nvidia rack were called NVL72.

Jensen Huang, co-founder and chief executive director of Nvidia Corp., performs during the Nvidia GPU Technology (GTC) conference in San Jose, California, USA, on Tuesday, March 18, 2025.

David Paul Morris | Bloomberg | Gets the image

Blackwell Ultra

Nvidia also announced new versions of her Blackwell chips, which she calls Blackwell Ultra.

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

Nvidia says it means that Cloud Providers can use Blackwell Ultra to offer AI Premium Services that feel time that allows them to earn 50 times more profits from new chips as a bunker generation in 2023.

Blackwell Ultra will come in version with two paired Nvidia ARM processor called GB300, and a version of the GPU called B300. It will also be given in versions with eight graphic processors in one server blade and a 72 Blackwell rack version.

Four -fourth cloud companies have deployed three times more than the number of chips blackwell in the form of a hopper chips, Nvidia said.

Depth

The Chinese Model Deepseek R1 may have frightened Nvidia investors when it was released in January but Nvidia accepted the software. The crush will use the model to focus on a few new products.

Many of the observers of the II said that the Deepseek model, which is reportedly, requires less chips than models made in the US, threatened the Nvidia business.

But Juan said that at the beginning of this year, Deepseek was actually a good sign for Nvidia. All because Deepseek uses a process called “reasoning”, which requires greater computing to provide users to the best answers.

New Blackwell Ultra chips are better for reflections, Nvidia said.

He developed his chips for a more effective conclusion, so when new reasoning models require more computing during deployment, Nvidia chips will be able to handle it.

“In the last 2 – 3 years there has been a major breakthrough, there has been fundamental progress in artificial intelligence. We call it an agentic ai,” Juan said. “It can think about how to answer or how to solve the problem.”