What is Liquid Cooling?
Liquid cooling uses a working fluid to remove heat from electronic components, offering higher efficiency than air cooling for high-density AI accelerators.
Liquid cooling is a thermal management method that transfers heat away from electronic components via a working fluid, typically water, dielectric oil, or a refrigerant. Unlike traditional air cooling, which relies on fans and heat sinks to convect heat into ambient air, liquid cooling can absorb large amounts of heat with a much smaller temperature gradient, enabling higher power density and lower fan noise.
The technology falls into several categories. Direct-to-chip cooling circulates liquid cold plates mounted directly on CPUs, GPUs, and memory modules. Immersion cooling submerges entire servers in a non-conductive dielectric fluid, eliminating the need for server fans and heat sinks. Single-phase immersion keeps the fluid in a liquid state; two-phase immersion allows the fluid to boil and condense, using latent heat for even greater efficiency. Rear-door heat exchangers sit at the server rack exhaust, cooling the hot air with liquid before it enters the room.
Liquid cooling has gained prominence with the rise of AI accelerators, such as NVIDIA H100 and AMD MI300X, which can exceed 700 W per GPU. Data center operators adopt liquid cooling to pack more compute into a rack without exceeding thermal limits or raising facility power budgets. Open standards, such as the Open Compute Project's Liquid Cooling specification, guide interoperability. Liquid cooling shifts heat rejection from computer room air handlers (CRAHs) to facility-level chillers or dry coolers, often reducing total power usage effectiveness (PUE).
Key facts
- Direct-to-chip liquid cooling uses cold plates attached to high-power components like GPUs and CPUs.
- Immersion cooling submerges servers in dielectric fluid, eliminating fans and enabling thermal densities above 100 kW per rack.
- Two-phase liquid cooling leverages latent heat of vaporization for higher efficiency at a given fluid temperature.
- The Open Compute Project publishes design specifications for liquid-cooled racks and coolant distribution units.
- AI accelerators with thermal design power (TDP) over 700 W require liquid cooling to stay within safe operating temperatures.
How it works in practice
Related terms
References
More in Data Centers
Carrier Hotel
A Carrier Hotel is a physical facility where multiple telecommunications carriers co-locate equipment and tenants can cross-connect directly to any carrier's network without using a third-party provider.
Carrier Neutral
A data center facility owned by an operator that does not sell network transit, allowing tenants to connect to multiple competing carriers and internet service providers.
Concurrent Maintainability
Concurrent maintainability is the ability to perform planned maintenance on any single component inside a datacenter without disrupting the IT load.
Data Center Tier Classification
The Uptime Institute's Data Center Tier Classification is a standard methodology for rating data center infrastructure based on redundancy, capacity, and availability, ranging from Tier I (basic) to Tier IV (fault-tolerant).
Edge Data Center
An edge data center is a small, distributed facility located close to end users to minimize latency and support real-time applications, often deployed as prefabricated units and operated remotely.
Free Cooling
Free cooling uses outside air or water that is cooler than the return temperature to reduce or eliminate the need for mechanical refrigeration in data center cooling systems.
Hot Aisle / Cold Aisle
Hot Aisle / Cold Aisle is a data center rack layout design that alternates rows of server intakes and exhausts to separate cool supply air from hot exhaust air, improving cooling efficiency.
Hyperscale Data Center
A hyperscale data center is a massive, single-tenant facility built by cloud, internet, or social-media giants to support tens of megawatts of IT load and hundreds of thousands of servers.
Internet Exchange Point
A physical infrastructure facility where multiple autonomous networks interconnect to exchange traffic directly, bypassing transit ISPs to reduce latency and cost.
kVA
kVA (kilovolt-ampere) is a unit of apparent power used to rate electrical equipment like UPSes and PDUs, equal to 1,000 volt-amperes, and differs from kilowatts when the power factor is not 1.0.