Cool IT! - Comparing Power and Throughput Performance of Liquid and Air Cooled AI Nodes

Explore a comprehensive 15-minute conference talk that evaluates the performance impact of different cooling techniques on AI infrastructure through direct comparison of liquid-cooled versus air-cooled server configurations. Learn how researchers conducted model training, fine-tuning, and stress tests to demonstrate that liquid cooling increases computational throughput while reducing node-level power consumption by up to 1.5 kW. Discover how liquid-cooled systems maintain GPU temperatures 15-25°C lower than air-cooled counterparts, enabling sustained performance without thermal throttling. Understand how efficiency gains scale with computational intensity, showing modest improvements for low-utilization tasks but substantial benefits during high-intensity workloads. Gain insights into measurable improvements in performance per watt, operational costs, and thermal stability that become increasingly critical for minimizing energy and water footprint as AI model complexity continues to grow.