Nvidia has introduced its new Vera Rubin chip system, designed to enhance AI data center operations. The system combines CPUs and GPUs, with one CPU for every two GPUs. The Vera Rubin NVL72 rack, a stack of chips in a liquid-cooled platform, is described as more 'plug-and-play' than previous products. OpenAI already uses one Vera Rubin rack, according to Nvidia executives. The system is part of Nvidia’s strategy to provide complete AI solutions, not just individual chips. It also features localized memory subsystems offering nearly three times the memory bandwidth of the previous Grace Blackwell super chip. Source: wired
Nvidia claims the Vera Rubin system will process 10 times as many tokens per watt as the Grace Blackwell super chip. The company also states its Vera CPU is faster at handling agentic AI tasks compared to rival CPUs from AMD and Intel, though the benchmarks used older generations of competitors' chips. The system is marketed as 'cable-free compute' and 'hot-swappable,' reducing installation time from hours to minutes. Nvidia’s Ian Buck emphasized the company's focus on innovating both GPUs and CPUs to stay competitive in Silicon Valley. The new chip system is fully liquid-cooled, which reduces energy use for cooling. Source: wired
Nvidia has been promoting Vera Rubin since its spring 2025 unveiling, with full production expected in the second half of 2025. Early customers include Microsoft, OpenAI, and Oracle. The company is particularly cautious about delays after previous-generation Blackwell chips overheated, leading to design changes and shipment delays. Nvidia’s marketing push coincides with AMD’s annual conference, where the company will showcase its Helios AI chip rack. Both companies are competing for contracts with AI hyperscalers and labs. Source: wired
