NVIDIA Blackwell platform slashes water use by 300x

Liquid cooling technology is emerging as a critical solution for managing the intense heat generated by AI computing systems, addressing a growing challenge in data center operations. As AI models continue to scale in size and complexity, traditional air cooling methods are becoming inadequate for dissipating heat from high-density server racks that now support over 135 kW compared to just 20 kW in previous generations. This shift toward liquid cooling represents not only a technical necessity but a strategic approach to improving energy efficiency and performance as AI infrastructure demands intensify.

The big picture: NVIDIA’s liquid-cooled systems are transforming data center efficiency by dramatically reducing reliance on traditional mechanical chillers while supporting advanced AI workloads.

• The NVIDIA GB200 NVL72 and GB300 NVL72 rack-scale systems are specifically designed to handle trillion-parameter large language model inference and AI reasoning models.
• These systems offer substantial improvements in energy management while maintaining the performance requirements for advanced AI applications.

Important stats: NVIDIA’s liquid-cooled architecture delivers remarkable efficiency gains that translate to both operational and environmental benefits.

• The systems provide 40x higher revenue potential and 30x higher throughput compared to traditional air-cooled architectures.
• Energy efficiency improves by 25x while water efficiency increases by an astounding 300x, addressing two critical resource constraints in data center operations.

Key details: The industry is exploring multiple cooling approaches to address the thermal challenges of high-density AI computing.

• Traditional mechanical chillers are being supplemented or replaced by alternative methods like evaporative cooling, dry coolers, and pumped refrigerant systems.
• Companies like Vertiv have developed reference architectures that reduce annual energy consumption by 25%.
• Schneider Electric’s solutions now support up to 132 kW per rack, while AWS reports increasing compute power by 12% while reducing energy consumption by up to 46%.

Where we go from here: NVIDIA is spearheading industry transformation through its COOLERCHIPS program, which aims to develop next-generation cooling systems.

• The program targets cost reductions of at least 5% and efficiency improvements of 20% over traditional air-cooled designs.
• These advancements will be crucial as AI models continue to grow in size and complexity, particularly for organizations deploying reasoning models at scale.