This week, NVIDIA is heavily promoting its new Vera Rubin chip system and has released the latest performance benchmark data for the combination of the GPU and CPU.
At a comprehensive technical seminar held last week at NVIDIA’s headquarters in Santa Clara, California, company executives presented to a small group of reporters a detailed focus on improvements in the chip system’s performance and energy efficiency. The core message conveyed by this release is that NVIDIA, long known for producing GPUs, is increasingly aiming to position itself as a supplier of CPUs capable of powering AI agents.
Although GPUs remain the primary hardware for companies to train and run AI models, as the industry shifts toward more complex, more agentic systems, demand for CPUs has increased as well. CPUs can coordinate data flow, network connections, and other software tasks. This may be one reason Nvidia has been actively positioning itself as a full AI systems supplier—not just a supplier of AI chips.
Vera Rubin is the successor to Nvidia’s hybrid superchip system Grace Blackwell, and a key part of the company’s recent AI industry development strategy. Its design philosophy is that one CPU is paired with every two GPUs. In a Vera Rubin NVL 72 superchip system, every 72 Rubin GPUs are paired with 36 Vera CPUs. Nvidia also sells Vera CPUs separately; the company has reportedly informed Chinese customers that the product may be available as early as August.
Nvidia executives emphasized that their new Vera Rubin NVL72 racks—stacking multiple chips on a liquid-cooled platform—are more “plug-and-play” than some of the company’s earlier products. During a visit to Nvidia’s data center lab in Silicon Valley, executives revealed that OpenAI is already using a Vera Rubin rack.
Nvidia CEO Huang Renxun was not in attendance last week at a symposium held in Santa Clara; at the time, he was in Japan announcing new partnerships with several Japanese companies to jointly develop AI technology for the robotics sector. The symposium brief was hosted by Ian Bark, Nvidia’s long-time vice president responsible for accelerated computing and CUDA software architect.
Bark told reporters, “We’re creating a roadmap aimed at rolling out new architecture—not only GPUs, but also CPUs. We’ll keep innovating, because in Silicon Valley, it’s innovate or die.”
A Nvidia spokesperson told WIRED that the event was held in Huang Renxun’s executive briefing center. On the desks nearby were stacks of bags of Taiwanese snacks—brought back by the CEO from Taipei International Computer Show (Computex, the annual large semiconductor trade show) recently.
Nvidia claims that the Vera Rubin NVL72 system delivers 10 times the processing power per watt as its Grace Blackwell superchip. The company says that, compared with competitors from AMD and Intel, its Vera CPUs are faster at handling agentic AI tasks (although the tests used to support these benchmarks appear to use slightly older generations of competing CPUs). The local memory subsystem on the new chip will also provide nearly three times the memory bandwidth of Blackwell. Given the current shortage of high-bandwidth memory, this is likely to be a highly attractive feature for many companies.
Nvidia says it has substantially reduced the number of cables needed to connect chips to racks in multi-rack server systems, to the point that the company markets the Vera Rubin system as “wireless cable computing” and “hot-pluggable.” That means, in theory, customers can shorten rack installation time from several hours to a few minutes, a point also jointly emphasized by Nvidia’s senior vice president of hardware engineering Andrew Bell and Bark. In addition, the new chip systems use 100% liquid cooling; because air cooling consumes more energy, this can reduce the energy required to cool the chips.
Since Nvidia announced that it would release the Vera Rubin chip in spring 2025, the company has been gradually revealing more details about the chip system while insisting that it will launch as scheduled. Huang Renxun has said multiple times that Vera Rubin is ramping mass production and will ship in the second half of this year. Early customers include Microsoft, OpenAI, and Oracle.
Earlier reports said that Nvidia’s previous-generation Blackwell chips would overheat when connected together in Nvidia’s custom server racks, forcing the company to make design changes and delay shipments. As a result, Nvidia is particularly sensitive to any hints of delays.
Nvidia’s marketing push for Vera Rubin came on the eve of its rival AMD’s annual meeting, at which AMD executives were expected to heavily promote their next-generation AI and data-center chips. On Sunday, AMD revealed more details about its Helios AI chip rack, designed to compete with Nvidia’s new products. Nvidia and AMD have been racing to win multi-year, large chip contracts from AI hyperscale data centers like Meta and Amazon, as well as AI labs such as OpenAI, Anthropic, and SpaceXAI.
Over the past two years, AMD has significantly grown its share in the data-center CPU market. The company has long been recognized as a pioneer in modern chip architectures used for x86 processors—x86 processors still account for the vast majority of data-center CPU revenue. By contrast, Nvidia builds its data-center CPUs on the ARM architecture, an architecture known for its energy efficiency.
Nvidia executives Bark and Hannah Kuutande, who is responsible for marketing Nvidia’s CPU products, both emphasized that the Vera Rubin processor does away with many of the chipset-style architectures used by modern processors, opting instead for a single-chip design. Kuutande noted that assembling multiple chipsets together would “seriously impair memory bandwidth and data transfer,” whereas Vera Rubin’s single-chip design allows data to be transmitted faster on a single integrated circuit.