top of page

Adiabatic and liquid cooling essential for GPU farms

  • Cedric KTORZA
  • Oct 15
  • 7 min read
ree

“L’adiabatique et le refroidissement liquide indispensables pour les fermes GPU.” In the age of AI-scale compute, adiabatic and liquid cooling have become the only credible path to safe, efficient, and scalable GPU farms.

When you train large models or run HPC workloads, heat density and power transients overwhelm legacy air-only designs. In this guide, we explain when and how to use adiabatic systems and liquid cooling, why hybrid architectures outperform single-mode cooling, and how Score Group integrates both—bridging energy, digital infrastructure, and new tech—to future‑proof your GPU clusters.

 

In brief

  • Adiabatic cooling slashes fan and chiller energy and enables higher supply temperatures; liquid cooling removes heat at the source for dense GPU racks.

  • Hybrid designs (direct‑to‑chip + adiabatic dry coolers) deliver high efficiency, resilience, and seasonal optimization with lower total cost of ownership over time.

  • Plan for water, not just watts: WUE, water quality, and hygiene protocols are first‑class design inputs.

  • Start with a measured roadmap: density mapping, CFD, lab PoC, then phased rollout with monitoring and leak‑safe controls.

  • Score Group aligns Energy, ITS, and Technology divisions to design, integrate, and operate end‑to‑end cooling for GPU farms.

 

Why GPUs change the cooling equation

High‑density GPU nodes concentrate far more power and thermal output than typical CPU servers. Air alone struggles to move enough heat across short distances without excessive fan power, noise, and hot spots. As clusters scale, operators face:

  • Heat flux at the chip exceeding what air heat sinks can remove efficiently.

  • Racks with tens of kilowatts of IT load, producing steep thermal gradients.

  • Transient “turbo” spikes that demand coolant mass and responsiveness.

Liquid cooling removes heat directly at the source; adiabatic systems reject heat efficiently outdoors with minimal mechanical compression. Together, they set a practical path to density, efficiency, and resilience. For environmental and operating envelopes, see ASHRAE’s guidance on liquid cooling for datacom equipment and allowable temperature ranges:

 

Adiabatic cooling fundamentals for data centers

Adiabatic (evaporative‑assisted) cooling pre‑conditions intake air or enhances dry coolers by adding moisture to reduce the air’s dry‑bulb temperature. Key advantages:

  • Efficiency: By leveraging wet‑bulb temperature, systems can avoid or reduce mechanical chilling, especially in dry climates.

  • Higher setpoints: Supports warmer supply air and water, enabling more “free cooling” hours annually.

  • Simplicity: Fewer compressors running means fewer failure points and lower maintenance in many designs.

What to consider:

  • Climate suitability: Performance depends on local wet‑bulb conditions and seasonal humidity.

  • Water management: Design for Water Usage Effectiveness (WUE), ensure filtration, scaling control, and hygiene protocols.

  • Air quality and filtration: Outdoor air integration demands robust filtration and corrosion management.

For metrics and best practices on PUE and WUE, see The Green Grid resources:

Additional engineering insights and real‑world case studies are available from neutral bodies:

 

Liquid cooling options for GPU racks

Liquid cooling brings a coolant close to or in contact with heat sources, capturing a large share of heat before it warms room air.

 

Direct‑to‑chip (D2C)

Cold plates on GPUs/CPUs connect via manifolds to a Coolant Distribution Unit (CDU). Benefits:

  • High heat capture at source, minimizing air recirculation.

  • Warmer water loops (e.g., 30–45°C supply in many designs) enable dry or adiabatic rejection without chillers.

  • Vendor‑supported for modern GPU platforms.

Design notes:

  • Secondary loop isolation (facility water system to IT loop) with heat exchangers for safety.

  • Leak detection, dripless quick‑disconnects, and service clearances.

For technology guidance and platform readiness:

 

Rear‑door heat exchangers (RDHx)

A passive or active coil replaces the rack’s rear door, removing heat as it exits the rack. Benefits:

  • Retrofit‑friendly and vendor‑agnostic for mixed IT.

  • Reduces room heat load, easing white‑space air handling.

 

Immersion cooling

IT is submerged in dielectric fluid (single‑phase) or immersed for two‑phase boiling. Benefits:

  • Exceptional heat capture and uniform temperatures.

  • Very high density potential.

Considerations:

  • Serviceability, fluid management, vendor ecosystem maturity, and facility adaptations.

For facility‑level design and integration patterns:

 

The power of hybrid: adiabatic rejection + liquid at the chip

A pragmatic blueprint for GPU farms:

  1. Capture heat at the source with D2C or RDHx to control hot spots and reduce fan energy.

  2. Reject heat via dry coolers enhanced with adiabatic assist to maximize compressor‑free hours.

  3. Operate at elevated water temperatures to enable free cooling and potential heat reuse.

  4. Retain air systems for memory, NICs, and ancillary components, tuned for lower airflow.

Plan for peak density and transient bursts, not fleet averages. Hybrid systems provide the thermal headroom you’ll need tomorrow, not just today.

Where feasible, channel warm water to heat reuse (district networks or building heat) using heat pumps if needed. European guidance on heat reuse effectiveness (HRE) and reporting can be found in EN 50600 and EU initiatives:

 

Design KPIs and governance

Measure what matters:

  • PUE (Power Usage Effectiveness): Overall efficiency of the facility.

  • WUE (Water Usage Effectiveness): Annual water use per IT energy (L/kWh).

  • HRE (Heat Reuse Effectiveness): Fraction of waste heat effectively reused.

  • Supply and return temperatures, approach temperatures, and ΔT across cold plates and coils.

  • Failure domain boundaries and redundancy (N+1/2N) for pumps, CDUs, and adiabatic modules.

A KPI‑driven design, aligned to ASHRAE classes and The Green Grid metrics, reduces risk and clarifies trade‑offs.

 

Risk, resilience, and compliance

  • Water stewardship: Ensure make‑up water availability, quality (conductivity, hardness), and blowdown strategy. Track WUE and seasonal patterns.

  • Hygiene and safety: Implement biocide dosing, materials compatibility, and maintenance routines per public health guidance for evaporative systems.

  • Controls and monitoring: Continuous leak detection, flow/pressure sensors, and predictive maintenance for pumps and valves.

  • Grid and energy policy: Warmer water loops enable higher chiller‑less hours, cutting peak demand and carbon intensity when paired with renewable supply.

For operations and resilience patterns:

 

Implementation roadmap with Score Group

At Score Group, we unite Energy, ITS, and New Tech to deliver end‑to‑end outcomes:

  1. Assessment and modeling (Noor ITS): IT load/density mapping, CFD of white space, facility hydraulics, and failure mode analysis.

  2. Energy integration (Noor Energy): Selection of dry/adiabatic coolers, pumps, heat exchangers, smart GTB/GTC, and heat reuse opportunities.

  3. New tech enablement (Noor Technology): IoT sensors, telemetry, AI‑driven anomaly detection, and automation across cooling loops.

  4. Pilot and validation: Lab PoC and a production pilot lane to validate performance, safety, and service workflows.

  5. Phased rollout and optimization: Progressive densification with PUE/WUE/HRE reporting and continuous improvement.

To explore how we can help, visit Score Group.

 

Choosing the right GPU cooling: quick comparison

Cooling approach

Best use cases

Energy efficiency potential

Water considerations

Complexity & retrofit

Notes

Adiabatic air (indirect/direct)

Moderate–high density rooms; seasonal free cooling

High, especially in dry climates

Requires make‑up water, treatment, hygiene protocols

Moderate; integrates with existing air systems

Extends chiller‑less operation; watch outdoor air quality

Direct‑to‑chip liquid

High‑density GPU racks; new builds or targeted retrofits

Very high by removing heat at source

Closed loops minimize consumption; quality control critical

Moderate–high; CDUs, manifolds, leak detection

Warmer water enables dry/adiabatic rejection and heat reuse

Rear‑door heat exchangers

Mixed IT and incremental densification

High in rack‑level removal

Similar to D2C (closed loop)

Retrofit‑friendly; vendor‑agnostic

Reduces room load without full liquid to components

Immersion cooling

Extreme density, constrained footprints

Very high

Fluid management vs. water; vendor‑specific

High; facility changes and service model shifts

Excellent thermals; consider ecosystem maturity

 

Practical engineering tips

  • Treat water as a design input: size storage, redundancy, and treatment from the start.

  • Elevate setpoints prudently: warmer water and air supply unlock efficiency—stay within IT and ASHRAE guidance.

  • Segment risk: isolate facility and IT loops; include quick‑disconnects and drip trays; test leak responses.

  • Instrument everything: pressure, flow, ΔT, and humidity data enable predictive cooling and early anomaly detection.

  • Prepare for maintenance: design clearances, bypass lines, and hot‑swap procedures for CDUs and pumps.

 

FAQ

 

Are adiabatic systems viable in humid climates?

Yes, with caveats. Adiabatic effectiveness depends on the gap between dry‑bulb and wet‑bulb temperatures. In humid seasons, the psychrometric leverage shrinks, so you’ll rely more on dry operation or auxiliary chilling. Many designs switch modes seasonally: dry only, adiabatic assist, and mechanical backup when needed. A site‑specific analysis of hourly weather data is essential to model annual performance, WUE, and water treatment needs. Standards like ASHRAE guidance help define acceptable thermal envelopes and control strategies.

 

Is liquid cooling mandatory for GPU servers?

Not universally, but it becomes practical or necessary as densities rise and as you target higher efficiency. Air can support moderate densities with excellent airflow design, but direct‑to‑chip or rear‑door heat exchangers remove heat at the source, enabling warmer water loops, lower fan energy, and better thermal stability during transients. Most modern GPU platforms provide liquid‑ready options. Evaluate per‑rack loads, transient behavior, and total cost of ownership, then validate with a pilot before broad rollout.

 

How do I manage water usage and hygiene safely?

Start with WUE targets and design water treatment accordingly. Use filtration, anti‑scaling measures, and dosing regimes compatible with materials. Implement automated purging/blowdown to control concentration cycles. Monitor conductivity, pH, and biocide levels, and schedule inspections of pads and coils. For closed IT loops, maintain glycol concentration and corrosion inhibitors. Document procedures aligned with public‑health guidance for evaporative systems, and audit routinely. The Green Grid’s WUE and ASHRAE references provide useful frameworks.

 

Can I retrofit liquid cooling into an existing data center?

Often, yes. Rear‑door heat exchangers are retrofit‑friendly, and direct‑to‑chip can be introduced in high‑density pods with CDUs and secondary loops. You’ll need structural reviews (manifolds, piping), segregated leak domains, and updates to BMS/GTB controls. A phased approach—pilot rack, then pod—limits risk and informs final design. Computational fluid dynamics, hydraulic modeling, and live telemetry de‑risk integration. Many operators combine limited liquid deployment for GPU pods with improved air handling elsewhere.

 

What metrics should I track during and after deployment?

Track PUE, WUE, and if applicable Heat Reuse Effectiveness, alongside rack inlet temperatures, water supply/return, ΔT across cold plates, pump power, and CDU efficiency. Monitor humidity, approach temperatures on dry/adiabatic coolers, and fan speeds. Use alarms for leaks, pressure drops, and abnormal transients. Establish seasonal baselines and compare against design intent. Reference metrics from The Green Grid and guidance from ASHRAE and OCP to ensure apples‑to‑apples reporting across phases.

 

Key takeaways

  • Adiabatic and liquid cooling are the practical foundation for dense, AI‑scale GPU farms.

  • Hybrid architectures capture heat at the source and reject it efficiently with minimal compression.

  • Water stewardship and hygiene are as critical as watts and airflow.

  • Start small, validate with a pilot, then scale with telemetry‑driven refinement.

  • Align energy systems, IT infrastructure, and new tech for resilient performance and sustainability.

Ready to design or upgrade your GPU cooling? Discover how we can help at Score Group.

 
 
bottom of page