Access Official Technical PDF

Share Summary:LinkedInEmail

Liquid cooling is an increasingly relevant option for data centers running high-performance computing (HPC) workloads. As processor designs add cores and power per socket rises, systems can produce more heat in a concentrated space. This puts pressure on rack power delivery, cooling capacity and airflow management. Liquid cooling can help address those constraints, but it works best as part of a broader efficiency strategy that also considers facility design, workload distribution and measurement.

Why data center cooling is under pressure

More cores and transistors per processor socket can increase compute capacity, while gains in performance and energy efficiency per core are slowing. The result is a growing thermal challenge: CPUs and GPUs can draw substantial power, and their heat must be removed reliably. Higher power density can contribute to hot spots and increase demands on rack power feeds, distribution systems and cooling infrastructure.

Air cooling remains a familiar approach, but concentrated heat loads can make it harder to deliver cooling evenly. Overcooling some areas while hot spots persist elsewhere is an inefficient response. A useful first step is to understand the facility’s actual conditions—including IT load, power use, temperature and humidity—rather than relying on broad assumptions about cooling needs.

Liquid cooling options at different scales

Liquid cooling is not one single technology. Options can be applied inside a server, at the rack, or across a data center. The right approach depends on the equipment, heat load, existing infrastructure and operational requirements.

In-server cooling

Direct-to-chip, closed-loop systems transfer heat from components such as CPUs and GPUs into a liquid cooling circuit. This can support higher-power processors and denser server configurations, while reducing reliance on air moving through the server. The source material also identifies actively pumped and passive in-server approaches, including heat pipes and solid-conduction designs.

Passive closed-loop systems use no active pumping components, while heat pipes and solid conduction transfer heat through physical materials. These approaches can avoid some active-part failure modes; fully sealed systems are also intended to avoid leakage. Their suitability depends on the particular server design and cooling requirement.

In-rack cooling

Rack-level solutions include rear-door heat exchangers, liquid-assisted air cooling and rack-mounted cooling distribution units (CDUs). Rear-door heat exchangers remove heat at the rack exhaust and can help contain hot air, reduce hot spots and limit unnecessary cooling. They may also be suitable for retrofitting existing racks, although space and installation requirements still need consideration.

Liquid-assisted air cooling captures heat from server cooling loops and transfers it through a liquid-to-air heat exchanger. A liquid-to-liquid CDU, by contrast, transfers heat from the rack loop to another liquid circuit. These arrangements can support higher rack heat loads, but their capacity and the proportion of server heat captured depend on the system configuration.

Immersion cooling

Immersion cooling places IT equipment in a cooling fluid. Single-phase and two-phase approaches are distinct designs; two-phase systems involve vapor management and can present servicing challenges. The whitepaper describes potential operational benefits such as reducing or eliminating server fans, lowering noise and reducing the number of fan-related failure points. These outcomes should be assessed against the equipment, enclosure and maintenance model involved.

What the reported HPC results illustrate

A project comparison in the whitepaper examined air-cooled and liquid-cooled systems in an OpenStack HPC cluster. Under CPU LINPACK testing, the listed CPU temperatures were 73°C and 76°C for the two air-cooled CPUs, compared with 52°C and 55°C for the liquid-cooled configuration. In GPU LINPACK testing, reported performance was 18.44 TFLOPS for air cooling and 19.07 TFLOPS for liquid cooling.

The same comparison lists GPU power at 300 W typical in both configurations, with maximum values of 375 W for air cooling and 379 W for liquid cooling. Reported GPU temperatures were 50–60°C typical and 65°C maximum with air cooling, versus 40°C typical and 49°C maximum with liquid cooling. These figures describe the particular systems and test conditions presented; they should not be treated as universal performance guarantees.

Connect cooling choices to facility efficiency

Power Usage Effectiveness (PUE) is a facility-level measure: total facility energy divided by IT equipment energy. A PUE of 1 represents the ideal in which facility energy is equal to IT energy. Liquid cooling can affect the energy used to remove heat, but it does not by itself determine a data center’s PUE. Power conversion, air management and other facility systems also matter.

Practical efficiency work can include:

  • Measure continuously: Track power, PUE, IT load, temperature and humidity to identify where energy and cooling capacity are being used.
  • Manage airflow: Use containment, barriers, blanking plates and side panels to reduce unwanted mixing of hot and cold air.
  • Match cooling to the load: Review tile layouts, cold-air outlets and equipment placement in relation to IT demand.
  • Assess operating conditions: Consider whether temperature and humidity ranges can be adjusted within appropriate equipment limits.
  • Review facility systems: Evaluate free cooling where climate permits, efficient UPS equipment and opportunities to reduce power-conversion losses.
  • Model before changing: Computational fluid dynamics (CFD) can help assess airflow and cooling designs before implementation.

Choose a cooling strategy, not just a technology

For plant, operations and infrastructure teams, the practical question is not simply whether liquid cooling is more efficient than air cooling. It is where heat is being generated, how it moves through the facility, and which cooling approach fits the equipment and operating model. In-server, in-rack and immersion solutions address different parts of that problem and bring different design and service considerations.

Liquid cooling can help support higher-density HPC systems and manage heat closer to its source. To turn that potential into facility-level gains, pair technology selection with careful data center design, load distribution and ongoing measurement. That combination gives decision-makers a clearer basis for balancing performance, energy use, reliability and operational needs.

Access Official Technical PDF

About The Author