Liquid cooling in compute engines: advantages and testing
Compare the advantages of direct-to-chip and immersion liquid cooling in compute engines, and learn how to test systems for thermal load and reliability.
In this article10
- Liquid cooling in compute engines: advantages and testing
- High-density compute engines: the limits of air cooling
- Energy inputs and grid externalities
- Direct-to-chip vs immersion cooling: advantages compared
- Energy efficiency and power usage advantages
- Criteria: how to test a liquid cooling system
- Reliability testing: how often does liquid cooling fail?
- Evaluating liquid cooling providers and system deployment
- Frequently asked questions
- Sources

Liquid cooling in compute engines: advantages and testing
High-density compute engines: the limits of air cooling
Modern silicon architectures used in artificial intelligence clusters push thermal dissipation beyond the heat transfer capacities of forced-air systems. High-density AI accelerators and central processing units exceed thermal design power thresholds that traditional computer room air conditioning units can clear efficiently. As power density per rack escalates from historical baselines of 5 kW to 10 kW up to 40 kW and higher, the physical volume of air required to cool individual sockets creates unmanageable airflow velocities and parasitic fan energy consumption.
| Parameter | Air Cooling Baseline | Direct-to-Chip Liquid Cooling | Two-Phase Immersion Cooling |
|---|---|---|---|
| Heat transfer medium | Air | Water / Glycol mixture | Dielectric fluid |
| Volumetric heat capacity ratio | 1x | ~3,500x | ~1,600x |
| Typical rack density | No universal ceiling; around 40.8 kW per NVIDIA's DGX H100 air-cooling guidance | Approx. 40 kW - 100 kW | Approx. 100 kW - 250 kW |
| PUE cooling overhead | 1.30 - 1.50 | 1.05 - 1.15 | 1.02 - 1.05 |
The acceleration of artificial intelligence workloads has fundamentally transformed the thermal density of modern compute engines. High-performance accelerators and server processors have reached thermal design power (TDP) thresholds that exceed the physical heat rejection capabilities of air cooling. Open Compute Project guidance for accelerator platforms describes cold plate loops sized for 700 W accelerator modules today, with demand anticipated for modules of up to 1000 W within a few hardware generations[7]. At those socket power levels, traditional air-based heatsinks and forced-air fan arrays encounter physical boundaries in heat transfer rates.
Energy inputs and grid externalities
Global datacenter electricity consumption reached approximately 415 TWh in 2024[12], about 1.5%[1] of total worldwide power demand, according to International Energy Agency analysis[1]. After the servers themselves, cooling and environmental control is the next-largest consumer of electricity inside those facilities[2]. That makes the thermal architecture, not only the silicon, a decisive factor in facility energy performance. Figures in this article were retrieved from the primary sources listed under Sources and are rechecked whenever those agencies and operators publish updated data.
European transmission system operators operate constrained high-voltage networks that dictate where new large-scale compute facilities can connect. Amprion manages an 11,000 km high-voltage grid spanning seven German federal states from the North Sea to the Alps[4], while TenneT oversees more than 26,000 km of high-voltage connections across Germany and the Netherlands[5]. The physical capacity limits of these networks force long queue times for grid interconnections in central European transmission corridors.
Network operators like 50Hertz are deploying capital to expand transmission lines and sub-stations, securing financing such as a EUR 500 million corporate bond dedicated to grid expansion required for energy transition demand[6]. Integrating high-efficiency liquid cooling systems reduces parasitic site power, allowing facility planners to maximize compute throughput within fixed grid allocation limits. Evaluating site energy infrastructure constraints early ensures that new compute deployments do not exceed available connection capacities.
Direct-to-chip vs immersion cooling: advantages compared
Choosing a liquid cooling architecture depends on compute density, facility layout, and retrofitting constraints. Direct-to-chip liquid cooling uses closed-loop cold plates attached directly to central processing units and graphics processing units. Coolant flows through microconvective channels within the cold plate to capture heat directly from the integrated heat spreader, while remaining rack components rely on low-velocity secondary airflow or rear-door heat exchangers.
Immersion cooling submerges entire server assemblies in tank-enclosed dielectric fluids. Single-phase immersion circulates non-conductive liquid through external heat exchangers via pumps, whereas two-phase immersion relies on low-boiling-point dielectric liquids that vaporize on hot component surfaces, condense on top-mounted coils, and drip back into the fluid bath. Two-phase systems achieve exceptional thermal transfer and remove the need for fans, cold plates, or custom server tubing assemblies.
| Evaluation Criterion | Direct-to-Chip (DLC) | Single-Phase Immersion | Two-Phase Immersion |
|---|---|---|---|
| Primary heat transfer mechanism | Conduction to microchanneled cold plate | Convection in dielectric bath | Phase change (boiling and condensation) |
| Target heat density per rack | 40 kW - 100 kW | 60 kW - 100 kW | 100 kW - 250+ kW |
| Facility retrofit complexity | Moderate (uses standard rack layouts) | High (requires horizontal tank floor placement) | High (requires sealed pressure-rated tanks) |
| Fluid maintenance requirement | Closed loop water/glycol monitoring | Dielectric oil filtration and sampling | Vapor capture and seal integrity inspection |
Enterprise operators evaluating NVIDIA AI chip pricing and hardware investments select direct-to-chip systems when retrofitting existing air-cooled raised-floor data centers. Direct-to-chip architectures preserve standard vertical rack form factors and maintenance workflows. Conversely, greenfield deployments designed for extreme compute densities benefit from immersion tanks, which eliminate internal fans completely and reduce building floor space requirements by up to 60%.
Energy efficiency and power usage advantages
Power Usage Effectiveness (PUE) measures facility energy efficiency by dividing total facility power by IT equipment power. Traditional air-cooled data centres frequently operate at PUE ratios between 1.30 and 1.50, meaning that 30 to 50 watts of overhead power are consumed for every 100 watts delivered to IT loads. Implementing liquid cooling lowers facility PUE, but the larger saving sits inside the server: in a Vertiv and NVIDIA analysis, moving roughly three quarters of the load to direct-to-chip cooling cut server fan power by 80% and IT power by 7% against an all-air baseline[8], lowering overall operational costs for high-density energy infrastructure.
High-density clusters operating modern AI accelerators generate concentrated heat that DLC and immersion architectures remove at lower operational delta-T temperatures. Direct-to-chip systems circulate liquid coolant through microchannel cold plates mounted on high-power chips (CPUs and GPUs), but they cannot absorb the whole load: in the most liquid-heavy case of the Vertiv and NVIDIA study, 74.9% of the load was cooled by liquid and the remaining 25.1% by air[8]. Immersion cooling instead submerges complete server chassis in a dielectric fluid, removing the need for server fans in the wetted portion of the system.
Higher heat transfer coefficients allow liquid systems to operate with higher supply fluid temperatures. Warm-water liquid cooling eliminates the need for mechanical chillers in standard European climate conditions, permitting year-round dry-cooling or evaporative heat rejection and driving annual facility PUE down toward 1.05 to 1.10.
Liquid cooling systems also facilitate waste heat capture. Higher fluid return temperatures enable facility operators to export thermal energy to district heating networks or industrial processes, supporting compliance with EU Energy Efficiency Directive disclosures.
Criteria: how to test a liquid cooling system
Verifying a liquid cooling system before production deployment requires systematic testing against published engineering standards. Procurement and engineering teams evaluate hydraulic balance, heat transfer efficiency, and power consumption under controlled baseline conditions. Establishing clear testing criteria ensures the cooling loop maintains thermal stability under peak computational stress without exceeding fluid operating parameters.
- Hydraulic pressure and flow testing: Measure flow rates (L/min) and pressure differential (kPa) across all secondary distribution loops under variable pump speeds.
- Thermal resistance evaluation: Validate junction-to-fluid thermal resistance (°C/W) against published methodology and manufacturer thresholds.
- ASHRAE thermal class alignment: Confirm supply water temperatures match target ASHRAE liquid-cooling temperature classes (W17 through W45 and W+).
- Energy efficiency baseline: Calculate CDU pump electrical input relative to total thermal kilowatts rejected to determine internal power overhead.
Operational testing relies on standardized thermal baselines established by industry standards. The fifth edition of the ASHRAE TC 9.9 thermal guidelines names its facility water classes after the maximum allowable supply temperature, giving classes W17, W27, W32, W40, W45 and W+ (above 45°C)[11]. Testing procedures must confirm that heat rejection components maintain steady-state junction temperatures under continuous maximum load within the designated water temperature class.
Independent research organizations provide specialized infrastructure for high-capacity thermal testing. For example, the Southwest Research Institute operates its Large-Scale Heat Exchanger Test Facility (LS-HXTF), configured with 5 MW of heating and cooling capacity to simulate actual data center flow conditions[10]. Facilities operating at this scale allow engineers to test full-scale coolant distribution units, heat exchangers, and high-flow pump skids under stress prior to data center deployment.
Reliability testing: how often does liquid cooling fail?
Industry design guidance from the Open Compute Project (OCP) sets reliability expectations for cold plate loops and chassis-level assemblies, defining both a target annual failure rate and a minimum mean time between failures at elevated coolant supply temperatures. Primary risks include micro-leaks at connection points, pump impeller degradation, and coolant contamination causing channel clogging inside micro-channel cold plates.
| Failure Mode | Primary Cause | Predictive Testing Method | Operational Mitigation |
|---|---|---|---|
| Quick-disconnect coupling leak | O-ring wear or improper latching | Helium mass spectrometry leak detection | Use drip-free flush-face quick disconnects |
| Galvanic corrosion | Incompatible metals in fluid loop | Electrical conductivity and pH monitoring | Maintain closed dielectric or copper-only loops |
| Biological growth / fouling | Untreated water-glycol mixture | Fluid sampling and biocidal analysis | Implement continuous inline filtration and UV dosing |
| Pump mechanical failure | Bearing wear or dry running | Vibration analysis and flow differential tracking | Deploy redundant N+1 pump skids in CDU |
Because liquid coolants possess high heat absorption capacity, loss of coolant flow leads to rapid thermal expansion at the silicon die. When a pump fails or a flow restriction occurs, chip temperatures can rise at rates exceeding 10°C per second. Mitigating downtime requires automated telemetry integration that triggers hardware-level power capping or immediate shutdown before critical silicon temperature limits are exceeded.
Evaluating liquid cooling providers and system deployment
A rigorous evaluation framework focuses on fluid circuit separation, leak detection capabilities, and hardware protection commitments. The most reliable comparisons come from published engineering criteria, such as OCP cold plate loop guidance and ASHRAE facility water classes, applied consistently to every shortlisted vendor architecture.
Comparing European AI infrastructure providers works best when every shortlisted architecture is scored against the same published documents. ASHRAE TC 9.9 supplies the facility water class definitions, from W17 through W45 and W+, that fix the supply temperature a design is rated for[11], while Open Compute Project cold plate guidance supplies the reliability and failure-rate targets. Asking each vendor which class its loops are commissioned to, and which OCP targets its cold plates and CDUs are held to, produces a like-for-like comparison rather than a set of unverifiable claims.
- Redundancy Architecture: Verify that coolant distribution units incorporate dual-pump N+1 or 2N configurations with active failover controls and hot-swappable pump modules.
- Fluid Containment & Leak Sensing: Ensure secondary loops incorporate continuous rope-style leak detection cables under manifolds, automated isolation valves, and non-conductive dielectric or low-conductivity fluids.
- Thermal Demarcation & Facility Interface: Confirm clear isolation between the facility water system (FWS) and the technology cooling system (TCS) via plate-and-frame heat exchangers.
- Service Level Guarantees and Response Commitments: Evaluate vendor SLAs for component replacement, on-site emergency support, and explicit hardware damage warranties resulting from cooling fluid loss.
Frequently asked questions
How to test a liquid cooling system?
Testing a liquid cooling system involves validating coolant flow rates, pressure levels, and thermal performance under both steady loads and simulated utility outages. Operators typically follow thermal guidelines published by ASHRAE to ensure cold plates and pumps function correctly before deploying the hardware in production.
How often does liquid cooling fail?
Exact failure frequencies vary by vendor and design, but industry guidance prioritises high reliability: the Open Compute Project publishes target annual failure rates and MTBF expectations for cold plates and coolant loops. Routine maintenance of pumps and continuous leak detection are critical to meeting those benchmarks.
Why is liquid cooling more efficient than air cooling?
Liquid possesses a significantly higher heat capacity than air, enabling it to absorb and transport thermal energy away from processors more effectively. This reduces the energy demand of secondary environmental control systems, which can account for over 30% of total power consumption in less efficient, air-cooled enterprise facilities.
What redundancy is required for liquid cooling systems?
To ensure continuous uptime, liquid cooling deployments typically employ 2N or N+1 redundancy configurations. This means installing backup pumps, secondary heat exchangers, and independent cooling loops so that if one component fails, the secondary system can immediately handle the thermal load without impacting computational performance.
Sources
- In focus: Data centres energy.ec.europa.eu
- What we know about energy use at U.S. data centers amid the AI boom pewresearch.org
- amprion.net
- tennet.eu
- Grid developement 50hertz.com
- opencompute.org
- Quantifying the impact on PUE and energy consumption when introducing liquid cooling into an air-cooled data… vertiv.com
- Data Center Cooling Testing & Research | Southwest Research Institute swri.org
- ashrae.org
Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score
Related reading
All Energy →Liquid cooling vs air: Rack density and energy overhead
Compare air, direct-to-chip and immersion cooling for data centers. Learn how rack power density, PUE and operational complexity determine the right choice.

Liquid cooling servers: Methods, fluids and density limits
Evaluate whether liquid cooling is right for your servers by comparing rear-door, direct-to-chip and immersion methods against rack-density thresholds.

Does the RTX 5090 Need Liquid Cooling? Air vs AIO Options
Determine if the 575W RTX 5090 needs liquid cooling. Compare AIO against air cooling, and learn whether you need to refill or replace the closed-loop liquid.
