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 6 of a nine-part series.
The term ‘liquid cooling’ encompasses both conductive (cold plates, heat pipes) and immersion approaches. Image by Rawpixel/Freepik
The conversation then delved into the intricate objectives of ASHRAE’s Work Statement 1972, as outlined by Davidson. He noted the extensive scope of the research, encompassing failure analysis and system design for both hybrid air-liquid and fully liquid-cooled equipment, utilising both modeling and empirical data. A key aspect involves investigating the impact of power and thermal capacitance on IT throttling behavior, with the aim of normalisation based on factors like current weight and processor/server types. The research also seeks to quantify the effect of the Technology Cooling System (TCS) loop’s liquid inlet temperatures, aligning with ASHRAE’s newly defined ‘S’ classes, and the impact of the TCS liquid delta T on IT throttling time – an area where current data is limited.
Furthermore, the research will assess the influence of the liquid cooling percentage in hybrid servers on the rate of temperature rise and IT throttling time, acknowledging that most non-immersion systems are a blend of liquid and air cooling. The resilience of the TCS loop will also be examined by evaluating the impact of liquid flow rate and pump failure percentages on overall IT uptime, differentiating between complete and partial pump failures. A significant component of the project involves developing approximately 30 targeted tests, combining various parameters to generate practical design information for the industry. Finally, the research will revisit the energy impact of liquid cooling, aiming to provide data that can inform updates to ASHRAE Standard 90.4, which currently lacks specific guidance on liquid cooling efficiency. Davidson mentioned that the 90.4 committee has agreed to provide their air-cooling-based calculations for comparison with the liquid cooling research findings, acknowledging the inherent complexities in such a direct comparison.
Seter acknowledged the ambitious scope of the research and inquired about industry support and donations to expedite this crucial work. He reiterated the pressing need for this guidance, particularly given the rapid adoption of liquid cooling. Seter then circled back to the concept of hybrid liquid-air cooling, prompting Quirk and Demetriou to elaborate on its implications for infrastructure design, especially when transitioning from traditional air-cooled densities of around 10 megawatts.
Quirk reiterated that the term ‘liquid cooling’ encompasses both conductive (cold plates, heat pipes) and immersion approaches, with the current discussion focused on cold plate applications. He stressed that even with liquid-to-chip cooling, the cooling of other server components necessitates a hybrid approach. A common emerging rule of thumb, the 80/20 principle, suggests that in a liquid-cooled server, approximately 80% of the heat is rejected via the liquid loop from the chips, while the remaining 20% is still handled by air cooling other components. Therefore, a 50-kilowatt liquid-cooled rack still presents a 10-kilowatt air-cooling load. As rack densities increase, the air-cooling demand also rises proportionally, eventually posing limitations even in hybrid systems.
