Liquid cooling featured highly in the recent Data Center World event in Austin, Texas. There were workshops on the subject, a variety of demonstrations, and quite a few chats in the hallways.
A couple of takeaways:
- There is confusion around nomenclature. “Liquid” is meant by some to be immersion; others to be water; and still others to include the entire spectrum of technologies other than traditional air cooling. In my mind, liquid cooling should include any technology that uses a liquid (refrigerant, oil, or water) to cool the heat-generating components of a data center.
- There already seems to be a bit too much dogma entering into such an early-stage market. There are a few certainties: Liquid in any form is better than air. Air is an inefficient transport for heat and is butting up against the limits of its usefulness. There are four main types of liquid cooling that are emerging on the scene. Each has pros and cons. The first is immersion: bathing your servers in a bath of dielectric fluid or mineral oil, which comes in the single- or two-phase variety. The third is single-phase water direct-to-chip, which means pumping water directly over the chips through “cold plates.” The fourth, two-phase DTC, is optimizing a phase change (boiling) of a refrigerant over a chip in the form of an evaporator plate.
To decide which technology is right for your environment, a few questions need to be answered.
- Is there legacy infrastructure which you would like to utilize?
- Would you like enough headroom on heat removal capability to ensure that your investment can be amortized over several generations of CPUs and GPUs?
- Are you concerned about your servers having a conductive fluid (water) running through them,
given that even small leaks within a server will destroy the server? - Do you want to maintain and leverage already existing service protocols for your support teams?
Different users will have different answers to these questions. With any of these technologies, there will be substantial savings on energy and dollars used to cool your data center. So upfront costs, failure domains, continued amortization of legacy infrastructure all become deciding factors.
Before allowing dogma to seep into your teams’ planning and evaluation, it would be a good idea to start with a discussion and prioritization of these basic goals.