An ASHRAE podcast recently delved into a critical evolution within data centres: the increasing necessity of liquid cooling. Host Justin Seter guided a panel of industry experts – David Quirk, Dustin Demetriou and Tom Davidson – through the intricacies of this technology, driven by the insatiable demands of artificial intelligence (AI) and high-performance Graphics Processing Unit (GPU) applications. This is Part 8 of a nine-part series.
A business-specific approach to tracking performance per watt might be more valuable than a universal metric. Image by Jordan Harrison/Unstock
The conversation then turned to the fundamental trade-off between the increased processing power enabled by high-density liquid cooling and the potential for decreased energy efficiency due to lower operating temperatures. Seter recalled the industry’s earlier focus on raising data centre temperatures for energy savings and questioned how the current trend towards liquid cooling and lower temperatures aligns with this goal.
Quirk referenced a figure in ASHRAE’s liquid cooling book illustrating this inherent challenge, rooted in the physics of silicon-based processors and conductive heat transfer. As chip power density increases, the temperature of the TCS loop’s inlet condition must be lowered to facilitate sufficient heat transfer. This inverse relationship means that the safety factors discussed earlier will further compound the need for lower temperatures. Quirk proposed reframing the concept of efficiency in this context, suggesting that the significantly higher processing power achieved in denser chips represents a greater ‘useful work’ output per unit of energy consumed. He acknowledged the difficulty for existing metrics like those in ASHRAE Standard 90.4 to fully capture this increased computational output.
Demetriou concurred, emphasising the need for metrics that reflect the ‘useful work’ achieved per unit of energy. He cautioned against solely focusing on traditional energy efficiency metrics at the infrastructure level, as seen in the earlier debates about raising air temperatures, where savings in mechanical cooling could be offset by increased server fan power. He also extended this challenge to water usage, advocating for new industry-wide metrics, perhaps similar to renewable energy credits, to promote sustainable water resource management alongside water-efficient cooling technologies.
Seter highlighted the relative simplicity of traditional metrics like PUE compared to the complexity of incorporating ‘useful compute’ as a denominator. He noted the industry’s ongoing efforts to define and measure this elusive metric, referencing initiatives like ENERGY STAR for servers and their server energy efficiency rating tools. However, he acknowledged the limitations of benchmarks and the difficulty in creating universally applicable metrics across diverse workloads and hardware. Demetriou echoed this sentiment, stressing that the definition of “useful work” can vary significantly between different businesses and organisations, suggesting that a business-specific approach to tracking performance per watt might be more valuable than a universal metric.
