▲ Nvidia
Nvidia, the world's leading AI semiconductor company, has begun supplying its next-generation AI platform, Vera Rubin (VR), the successor to Grace Blackwell (GB), in a move to solidify its market leadership.
Nvidia announced on July 21 (local time) that its Vera Rubin NVL72 AI racks are currently being supplied to and operated by Google Cloud, Microsoft Azure, Oracle Cloud, and CoreWeave.
Vera Rubin NVL72 is an AI supercomputer that integrates seven new chips, including Nvidia's Rubin graphics processing unit (GPU) and Vera central processing unit (CPU), into a single system housed within one cabinet.
Nvidia explained that it has completed its largest and most advanced supply chain to date, encompassing 350 factories across 30 countries, to meet the surging demand for computing power.
Performance tests conducted by CoreWeave, which installed the Vera Rubin platform, using the DeepSeek R1 model showed that the new platform delivers 10 times the performance per megawatt (MW) compared to its predecessor, Grace Blackwell.
The company also emphasized that its design, which allows for liquid cooling fluid intake at 45°C (113°F), can save millions of gallons (thousands of tons) of water annually per megawatt.
In terms of networking, the system incorporates the Spectrum-6 Ethernet switch system, which doubles the capacity of the previous generation to transmit 102.4 terabits per second (Tbps), enabling massive AI factories to operate as a single system.
Ian Buck, vice president at Nvidia, stated during a briefing at the company's headquarters in California that "we are in full production" and that "the system is up and running at all major customers," as reported by Bloomberg.
Nvidia's active promotion of its product performance is interpreted as a response to the intensifying competition in the AI chip market.
AMD recently announced that it is supplying its Helios AI rack, which competes with Vera Rubin, to major customers including Microsoft.
OpenAI and Meta, which were major customers of Nvidia, have recently unveiled their own AI chips, and reports suggest that Anthropic is also considering developing its own chips.
Meanwhile, Nvidia is also pursuing Intel and AMD in the CPU market.
During the briefing, Nvidia claimed that its Vera CPU runs the Python coding language up to 1.8 times faster than AMD's Turin.
However, AMD is scheduled to release Venice, the next-generation successor to Turin, soon.
Although there have been claims that the launch of the Kaiber NVL144 rack architecture, which can house Nvidia's Rubin Ultra chips, has been delayed by more than 12 months to 2028, Nvidia has denied these reports.
(Photo: AP, Yonhap News)
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.