Nvidia Launches Next-Gen AI Chips, Rubin Platform Enters Production

Nvidia Chief Executive Officer Jensen Huang announced the company's next generation of chips is in "full production," designed to deliver five times the artificial-intelligence computing power of previous chips for chatbots and other AI applications. These chips are expected to debut later this year.

Credit: NVIDIA

Huang revealed new details about these chips at the Consumer Electronics Show in Las Vegas. Nvidia executives stated the chips are already being tested by AI firms in the company's laboratories.

The Vera Rubin platform, composed of six individual Nvidia chips, is slated for release later this year. Its flagship server will contain 72 of the company's graphics units and 36 of its new central processors.

Huang demonstrated how these components can be combined into "pods" with over 1,000 Rubin chips. He explained that this configuration could improve the efficiency of generating "tokens," the fundamental unit of AI systems, by 10 times.

To achieve these performance results, the Rubin chips utilise a proprietary data type. Nvidia hopes the wider industry will adopt this new data standard to leverage the full capabilities.

Huang stated that this proprietary data allowed the company to deliver a "gigantic step up in performance," even with only 1.6 times the number of transistors compared to previous generations.

Nvidia faces increased competition in the AI market from traditional rivals, such as Advanced Micro Devices, and customers like Alphabet's Google. These competitors are actively involved in delivering AI model outputs to hundreds of millions of users.

A significant portion of Huang’s address focused on the new chips’ efficacy for this task, including the addition of a new "context memory storage" layer. This technology aims to help chatbots provide quicker responses to lengthy questions and conversations.

The company also introduced a new generation of networking switches featuring a "co-packaged optics" connection. This technology, crucial for linking thousands of machines, competes with offerings from Broadcom and Cisco Systems.

CoreWeave will be among the first to receive the new Vera Rubin systems. Microsoft, Oracle, Amazon, and Alphabet are also expected to adopt these systems, expanding their AI infrastructure.

In other announcements, Huang highlighted new Alpamayo software designed to assist self-driving cars in making path decisions and creating a paper trail for engineers. This software, along with its training data, will be released more widely for automakers to evaluate.

Credit: NVIDIA

Nvidia is open-sourcing both the models and the data used to train them. Huang noted this practice is intended to build trust in how the models were developed.

Nvidia recently acquired talent and chip technology from the startup Groq, including executives instrumental in developing Alphabet's Google AI chips. While Google is a major Nvidia customer, its own chips have emerged as a significant threat.

Google is working closely with Meta Platforms and other organisations to challenge Nvidia's stronghold in the AI market. Huang stated that the Groq deal "won't affect our core business" but could lead to new products that expand Nvidia's lineup.

Nvidia is keen to demonstrate that its latest products outperform older chips, such as the H200. This chip, a predecessor to Nvidia's current Blackwell chip, is in high demand in China.

United States President Donald Trump has permitted the H200 chips to be shipped to China. This situation has caused alarm among China hawks across the U.S. political spectrum.

Huang noted strong demand for H200 chips in China. Chief Financial Officer Colette Kress confirmed Nvidia has applied for licenses to ship the chips but is awaiting approvals from the U.S. and other governments.

• Nvidia's next-generation AI chips, offering five times the computing power, are now in "full production" and are expected later this year.

• The Vera Rubin platform, incorporating a proprietary data type and aiming for a 10-fold efficiency increase in AI token generation, will debut later this year.

• Nvidia is expanding its product ecosystem with new networking switches, "context memory storage" for chatbots, and Alpamayo software for autonomous vehicles.

Source: REUTERS