Liquid Cooling PUE FAQ: Expert Answers to Technical & Deployment Questions

Overview & Thematic Scope

Liquid cooling has moved from a niche HPC solution to a mainstream requirement for AI-density racks exceeding 100 kW. The central engineering question for datacenter architects is no longer whether liquid cooling works, but how precisely it drives Power Usage Effectiveness (PUE) down to 1.15 or below. This FAQ addresses the technical mechanisms, thermal design choices, and operational factors that determine whether a liquid-cooled facility achieves that target, drawing on published research and deployed system data.

Frequently Asked Questions

Q1: How does liquid cooling achieve PUE below 1.15 compared to air cooling?

Liquid cooling achieves PUE below 1.15 by eliminating the energy-intensive air-handling chain and enabling higher coolant supply temperatures that unlock free cooling. In a cold-plate system, coolant at 30–45°C absorbs heat directly from chips, and a facility water loop at 15–25°C rejects that heat through dry coolers or cooling towers with minimal compressor operation . Because liquid captures heat at the source rather than conditioning an entire room, the non-IT overhead shrinks from roughly 40–50% in legacy air-cooled halls to about 13% or less, which corresponds directly to a PUE of approximately 1.15 .

Q2: What is the difference between cold plate and immersion cooling for hitting PUE 1.15?

Both cold plate and immersion architectures can reach PUE 1.15, but they differ in heat capture efficiency and infrastructure complexity. Cold plate cooling uses sealed cold plates attached to high-power components, circulating coolant through a CDU to a facility water system; published data shows cold plate systems at 15°C supply temperature achieving PUE 1.15, while warm-water variants at 40–47°C inlet reach the same target with greater free-cooling hours . Immersion cooling submerges entire servers in dielectric fluid, and deployed systems have reported PUE as low as 1.09 in full-scale installations, though fluid cost and maintenance access remain trade-offs .

Q3: Why does higher coolant supply temperature improve PUE in liquid-cooled datacenters?

Higher coolant supply temperature improves PUE by expanding the annual hours during which chillers can be bypassed in favor of free cooling. When the facility water loop runs at 15–25°C, conventional chillers operate only during peak ambient conditions; in colder seasons, dry coolers or plate heat exchangers handle the load alone, cutting compressor power . Research on supply temperature matrices confirms that elevating the secondary-side setpoint allows conventional chillers to be gradually replaced by cooling towers or dry coolers, reducing facility water system rated power by up to 75% .

Q4: What role does the CDU play in achieving and maintaining PUE below 1.15?

The CDU is the hydraulic and thermal interface that isolates the technology cooling loop from facility water, and its approach temperature directly affects achievable PUE. A CDU with a tight approach temperature transfers heat from the 30–45°C server loop to the 15–25°C facility loop with minimal temperature penalty, allowing the facility side to remain warm enough for free cooling . However, CDU internal pressure drops and pump power add to non-IT overhead; research notes that CDU presence imposes structural constraints on flow optimization, and eliminating it entirely has been explored as a way to reduce parasitic energy .

Q5: Can liquid cooling maintain PUE 1.15 under dynamic AI training loads?

Maintaining PUE 1.15 under AI load swings requires control systems that track thermal transients in seconds rather than minutes. AI workloads can shift from near-idle to full GPU utilization rapidly, and conservative fixed-speed cooling strategies waste pump and fan energy during low-load periods while risking thermal overshoot during spikes . Modern liquid cooling systems address this with variable-frequency chillers and AI-driven optimization algorithms that adjust coolant flow and supply temperature in 3–5 second response windows, holding temperature stability within ±0.5°C while minimizing parasitic power .

Q6: What are the main deployment trade-offs when targeting PUE 1.15 with liquid cooling?

The primary trade-offs are higher upfront capital cost, facility water availability, and maintenance complexity compared to air cooling. Stanford analysis notes that while warm-water liquid cooling achieves the lowest reported PUE at 1.15, the total cost of ownership of liquid-cooled systems is likely higher than air cooling due to CDU, plumbing, and coolant infrastructure . Water-based heat rejection also introduces water consumption and treatment requirements, though air-cooled chillers or hybrid dry coolers can mitigate this at the cost of some efficiency during peak summer conditions .

Q7: What PUE results have been demonstrated in real liquid-cooled datacenter deployments?

Real-world deployments have consistently demonstrated PUE at or below 1.15 with liquid cooling, and immersion sites have reported values approaching 1.09. Alibaba’s Renhe full-immersion datacenter achieved a minimum PUE of 1.09, with annual electricity savings of 70 million kWh . Cold-plate systems from multiple vendors are specified at PUE ≤1.15 for AI datacenter deployments, including ZTE’s M8000 air-liquid shared-source POD and Gree’s full-stack liquid cooling solutions delivered to China Unicom and other intelligent computing centers . These results confirm that 1.15 is an achievable, repeatable target rather than a laboratory best-case.