Rear-Door Heat Exchangers for GPU Data Centers: Scalable Retrofit Cooling

The Cooling Gap Between Air and Full Liquid

Traditional air cooling, whether through raised-floor CRAC units or hot-aisle/cold-aisle containment, works reliably up to roughly 15 to 20 kW per rack. Full direct-to-chip liquid cooling or immersion cooling can handle 80 to 200+ kW per rack but requires specialized plumbing, leak containment, and often purpose-built facilities.

Between these two extremes sits a practical gap: racks drawing 20 to 60 kW where air cooling struggles but full liquid cooling is either too expensive to retrofit or impractical for the facility. Rear-door heat exchangers (RDHx) fill this gap. They attach to the back of standard server racks and intercept hot exhaust air before it enters the room, removing a significant portion of the heat via chilled water flowing through a coil. No modifications to the servers. No liquid inside the rack. No immersion tanks.

For data center operators looking to support rising GPU power densities without a full facility rebuild, RDHx is often the fastest and least disruptive path forward.

How Rear-Door Heat Exchangers Work

An RDHx unit replaces or attaches to the rear door of a standard 19-inch rack. Inside the door is a heat exchanger coil, typically copper or aluminum tubing with fins, through which chilled water circulates. Hot air exhausted by the servers passes through this coil before exiting the rack. The chilled water absorbs the heat and carries it to the building's chilled water loop, cooling tower, or dedicated coolant distribution unit (CDU).

Passive RDHx

Passive units have no fans of their own. They rely entirely on the server fans to push hot exhaust air through the coil. This is the simplest configuration: no power connection to the door, no additional noise, no moving parts to fail. Passive RDHx units typically remove 15 to 30 kW per rack, depending on the coil surface area, chilled water temperature, water flow rate, and server fan speed.

Passive units work best when the servers already have high-airflow fans (common in GPU servers). The harder the server fans push, the more heat the passive coil captures. For racks where server fan output is modest, passive units may not intercept enough heat to keep exhaust temperatures within acceptable limits.

Active RDHx

Active units include integrated fans behind the heat exchanger coil. These fans actively pull air through the coil regardless of server fan output, providing consistent and higher heat removal capacity. Active RDHx units can handle 40 to 80 kW per rack, with some high-capacity models rated for up to 100 kW under optimal conditions.

Active units consume 200 to 800 watts of fan power and generate additional noise (typically 55 to 65 dBA at full speed). They also require a power connection, adding a cord to manage. However, for GPU racks in the 30 to 60 kW range, active RDHx often provides the best balance of cooling capacity, deployment speed, and cost.

RDHx vs. Other Cooling Technologies

Technology	Capacity per Rack	Retrofit Difficulty	Best For
Room-level air cooling	5 - 15 kW	Already in place	Traditional IT, low-density racks
In-row cooling units	15 - 40 kW	Moderate (floor space)	Medium density, available floor space
Passive RDHx	15 - 30 kW	Low (rack-mount + water)	Retrofit, no power to door needed
Active RDHx	40 - 80+ kW	Low (rack-mount + water + power)	GPU racks, retrofit, fast deployment
Direct-to-chip liquid	50 - 150+ kW	High (plumbing to each server)	New builds, highest density
Immersion cooling	80 - 200+ kW	High (tanks, fluid, re-engineering)	Extreme density, purpose-built

For a comprehensive analysis of all cooling approaches, see our data center cooling technology comparison. The key differentiator for RDHx is the combination of meaningful cooling capacity with minimal disruption to existing infrastructure.

Infrastructure Requirements

Deploying RDHx requires supporting infrastructure that many data centers already have or can add incrementally.

Chilled Water Supply

RDHx units need a chilled water supply and return piped to each rack. The water temperature and flow rate directly determine cooling capacity. Lower supply water temperatures enable more heat removal but may introduce condensation risk if below the dew point of the data center environment. Typical supply temperatures range from 7 to 18 degrees Celsius, with 12 to 15 degrees Celsius being a common sweet spot that balances cooling capacity against condensation management.

Flow rates vary by unit and load. A 40 kW active RDHx with 12-degree supply water typically requires 4 to 6 liters per second of chilled water flow. The facility's chilled water plant must have sufficient capacity to serve the additional load. For sites considering liquid cooling in hot climates like the UAE, the chilled water infrastructure for RDHx is typically a subset of what full direct-to-chip or immersion would require.

Piping and Distribution

Water piping must be routed to each RDHx-equipped rack. This typically means overhead or underfloor distribution manifolds with flexible hoses dropping to individual racks. Quick-disconnect fittings allow door removal for maintenance without draining the entire loop. Leak detection systems (drip trays, moisture sensors, flow monitoring) are essential wherever water runs near IT equipment.

Structural Considerations

A fully loaded RDHx door weighs 40 to 80 kg depending on model and size. The rack frame must support this weight on its rear hinge points. Most modern data center racks accommodate this without modification, but older or lighter-gauge racks may need reinforcement. Additionally, the RDHx door adds 10 to 20 cm of depth to the rack, which can be a constraint in facilities with tight row spacing.

When RDHx Is the Right Choice

Retrofitting Existing Facilities for AI

The most common RDHx use case is adding GPU server racks to a facility that was designed for traditional air-cooled IT. Rather than rebuilding the cooling plant or installing immersion tanks, the operator adds RDHx doors to the GPU racks and extends the existing chilled water infrastructure. The rest of the facility continues operating unchanged.

Phased GPU Deployment

Organizations deploying GPU infrastructure in phases benefit from the rack-level granularity of RDHx. Deploy 5 racks of GPU servers with RDHx in the first phase. If the workload grows, add more RDHx-equipped racks without redesigning the cooling architecture. This phased approach aligns cooling capital expenditure with actual demand rather than requiring a large upfront investment in cooling infrastructure for future growth.

Mixed-Density Environments

Many data centers run a mix of traditional low-density compute (5 to 10 kW/rack) alongside newer high-density GPU racks (30 to 60 kW/rack). RDHx allows high-density racks to coexist in the same hall as low-density racks without overloading the room-level cooling that serves the low-density equipment. This avoids the cost and disruption of segregating GPU workloads into a separate, purpose-built space.

Colocation Environments

Colocation providers supporting AI tenants can deploy RDHx on a per-cage or per-rack basis, allowing them to offer high-density colocation without converting entire halls to liquid cooling. Tenants deploying GPU servers get the cooling they need; tenants running standard IT are unaffected. Rax Data & Energy incorporates both RDHx and direct liquid cooling options across its colocation facilities to support deployments at any density.

When RDHx Is Not Enough

Rear-door heat exchangers have limits. Understanding those limits prevents deploying RDHx where a different technology is the correct solution.

• Racks above 80 kW. Next-generation GPU systems like the NVIDIA GB200 NVL72 can draw 100 to 120 kW per rack. At these densities, even high-capacity active RDHx units cannot reject enough heat through the door alone. Direct-to-chip liquid cooling, which removes heat at the silicon before it reaches the air, is the appropriate technology for racks in this range.
• Extreme ambient temperatures. In regions like the UAE where outdoor temperatures regularly exceed 45 degrees Celsius, the chilled water plant must work harder to deliver low supply temperatures. RDHx cooling capacity is directly proportional to the temperature difference between the hot exhaust air and the chilled water. If the chilled water supply cannot be maintained below 15 degrees Celsius, RDHx capacity is reduced. Facilities in hot climates should oversize the chilled water plant or consider supplementing RDHx with in-row cooling for the highest-density racks.
• Space-constrained racks. The 10 to 20 cm of additional depth that an RDHx door adds can be a problem in facilities with tight row spacing. If rear clearance is already limited, installing RDHx may require rearranging racks or reducing usable rack units, which has its own cost implications.

Installation and Operational Considerations

Deployment Timeline

One of the strongest arguments for RDHx is speed. Installing a rear-door heat exchanger on an existing rack takes hours, not weeks. The longest lead-time item is typically the chilled water piping, which can be pre-staged during normal maintenance windows. Once piping is in place, attaching the RDHx door, connecting the quick-disconnect hoses, and commissioning the unit is a same-day operation per rack.

Maintenance

RDHx units require periodic maintenance: coil cleaning (dust and debris accumulation reduces heat transfer), filter replacement on active units, inspection of hose connections and fittings, and verification of water quality to prevent corrosion and biological growth in the loop. Most operators include RDHx maintenance in their standard quarterly data center maintenance cycle.

Monitoring

Each RDHx unit should be monitored for supply and return water temperatures, flow rate, inlet and outlet air temperatures, and (for active units) fan speed and status. These metrics integrate into the facility's DCIM platform to provide real-time visibility into cooling performance per rack. Alerts on declining delta-T (temperature difference between supply and return water) can indicate coil fouling, reduced water flow, or an increase in rack heat load that is approaching the unit's capacity.

Cost and ROI

RDHx costs vary by manufacturer, capacity, and whether the unit is passive or active. As a general range:

• Passive RDHx: typically in the range of several thousand dollars per unit, excluding piping
• Active RDHx: typically higher, scaling with cooling capacity and fan configuration
• Chilled water piping: variable by facility layout, distance from chilled water plant, and number of racks

The ROI case for RDHx is strongest when compared to the alternative of building new, purpose-built high-density space. Converting an existing data hall to support GPU racks using RDHx can be accomplished at a fraction of the cost and timeline of constructing or retrofitting for full immersion cooling. For operators that need to support AI workloads within months rather than years, RDHx is often the only viable path that does not involve cloud GPU rental at premium rates.

For organizations evaluating the full spectrum of colocation pricing and cooling options, the total cost of ownership analysis should include not just the RDHx hardware but also the chilled water infrastructure, ongoing water treatment, maintenance labor, and the power consumed by active unit fans.

Frequently Asked Questions

How much heat can a rear-door heat exchanger remove?

Passive RDHx units typically handle 15 to 30 kW per rack. Active (fan-assisted) units can remove 40 to 80 kW per rack depending on chilled water temperature, flow rate, and coil size. Some high-capacity active models are rated up to 100 kW under optimal conditions.

Can RDHx cool GPU racks without additional room cooling?

For racks up to approximately 30 kW, a properly sized active RDHx can capture the majority of exhaust heat and may reduce the need for supplemental room cooling in that zone. For racks above 40 kW, existing room cooling should remain active to handle residual heat, ambient temperature, and airflow for non-rack components.

What is the difference between passive and active RDHx?

Passive units have no fans and rely on server exhaust to push air through the coil (15 to 30 kW capacity, no power needed). Active units include integrated fans for consistent, higher-capacity heat removal (40 to 80+ kW, requires power connection, adds some noise).