NVIDIA Vera Rubin NVL72 Delivers 10x Energy Efficiency Boost Running DeepSeek R1
CoreWeave's testing of NVIDIA's next-generation Vera Rubin NVL72 platform revealed a 10-fold increase in token throughput per megawatt compared to the Blackwell-based GB200 NVL72 when running the DeepSeek R1 reasoning model. NVIDIA also announced that it is accelerating the mass production of the Vera Rubin NVL72, which is already running within partner infrastructures like CoreWeave, Google Cloud, Microsoft Azure, and Oracle Cloud. This massive leap in energy efficiency addresses one of the AI industry's biggest bottlenecks: power consumption. By delivering ten times the throughput per megawatt, the Vera Rubin architecture could significantly lower operational costs and environmental impact for running advanced reasoning models at scale. The benchmark focused on the relationship between token throughput per megawatt and interactivity (tokens per second per user) to ensure realistic user response times. CoreWeave completed the industry's first power-on and validation of the end-to-end Vera Rubin NVL72 rack-scale system in early June.
## BACKGROUND
The NVIDIA Vera Rubin NVL72 is a liquid-cooled rack-scale supercomputer that integrates 72 Rubin GPUs and 36 Vera CPUs interconnected via NVLink 6. DeepSeek R1 is a highly cost-effective, open-weight reasoning model developed by Chinese AI firm DeepSeek, which uses Mixture of Experts (MoE) architecture to achieve performance comparable to proprietary models like OpenAI's GPT-4.