NVIDIA reveals the latest progress of Vera Rubin, the chip has been delivered to Open AI and other companies
Nvidia revealed that Vera CPUs were delivered to customers in June, including OpenAI, Anthropic and SpaceX. In addition, NVIDIA revealed more information about the VeraRubin platform through a blog on July 21, local time. The VeraRubin platform includes 7 chips and 5 racks. NVIDIA said that the VeraRubin platform is accelerating towards gigawatt-scale deployment.
Nvidia revealed that Vera CPUs were delivered to customers in June, including OpenAI, Anthropic and SpaceX. In addition, NVIDIA revealed more information about the Vera Rubin platform through a blog on July 21, local time. The Vera Rubin platform includes 7 chips and 5 racks. NVIDIA said that the Vera Rubin platform is accelerating towards gigawatt-scale deployment. Vera Rubin has a supply chain system covering 350 manufacturing bases in 30 countries around the world. It is the largest and most mature rack-level supply chain in history. Nvidia said that the Vera Rubin platform has been optimized from chip to architecture, achieving the highest performance per watt and the lowest token (word unit) cost. CoreWeave's test results on DeepSeek-R1 show that the Vera Rubin platform's throughput per megawatt is 10 times that of the previous generation Grace Blackwell NVL72, which is a very important indicator for AI factories with limited power supply. Regarding the different products of the Vera platform, Nvidia stated that the Vera CPU is specially designed for the era of agents. Its customized Olympus core single-threaded performance has been increased by 2 times, and the inter-core bandwidth has been increased by 3 times. Compared with similar small chip designs, this chip is the most efficient single-threaded CPU when processing key agent workloads. In terms of network, the sixth generation NVLink can provide more than 2 times the throughput and reduce latency by 3 times under complex workloads. In terms of horizontal expansion, Spectrum-X Ethernet combines the 102.4T Spectrum-6 switch system, 1.6T ConnetX-9 SuperNIC network card, etc., making the RDMA bandwidth 1.6 times that of Ethernet products on the market. CoreWeave, Microsoft, SpaceX and Tesla have already taken the lead in introducing Spectrum-6. Production of the Vera Rubin NVL72 is accelerating. Nvidia said the relevant racks are already running on partner CoreWeave, Google Cloud, Microsoft Azure and Oracle cloud infrastructure. There are no cables, fans or hoses in the Vera Rubin NVL72 chassis, computing module assembly time has been reduced from hours to one minute, and a 45°C liquid cooling inlet design enables dry cooling operation without a chiller. NVIDIA said that for new AI factories, this high-temperature dry cooling combined with a closed-loop liquid cooling system can save hundreds of gallons of water per megawatt per year. In addition, Nvidia said that the Vera Rubin platform is the basis for Microsoft's cooperation with French AI company Mistral. The two companies plan to invest billions of dollars in building computing infrastructure in Europe. Mistral is deploying thousands of the latest Vera Rubin GPUs to increase its GPU capacity to improve customers' AI computing resource availability. NVIDIA CEO Jen-Hsun Huang also mentioned the Vera Rubin platform at last month's shareholder meeting, saying that Vera Rubin is built for intelligent agents. Vera Rubin has been fully put into production. Every major model developer, public cloud, AI cloud and hyperscale cloud vendors are preparing to build based on Vera Rubin. Vera CPU has begun to receive orders. NVIDIA previously estimated that Vera CPU would open up a new market of US$200 billion for the company.