In-Row CDU vs. In-Rack CDU: A Comprehensive Analysis | 2026

The technological trajectory of the modern data center is increasingly defined by a thermal crisis that traditional cooling methodologies are fundamentally unequipped to resolve. As artificial intelligence (AI), high-performance computing (HPC), and complex machine learning clusters become the standard for enterprise and hyperscale operations, the resulting power densities have shattered the 20 kW to 30 kW threshold where air cooling remains economically and physically viable. At the center of this transition toward liquid-based thermal management is the Coolant Distribution Unit (CDU), a sophisticated mechanical interface that serves as the heart of the liquid cooling ecosystem.

 

The selection between in-row and in-rack CDU architectures represents a critical decision point for data center architects, influencing not only immediate thermal performance but also long-term scalability, maintenance overhead, and total cost of ownership.

 

 

The Thermodynamic Imperative for Liquid-Based Cooling

The transition from air to liquid cooling is driven by the fundamental physical properties of the cooling media. Air is an inherently poor thermal conductor with a low volumetric heat capacity, requiring massive volumes and high-velocity movement to reject the heat generated by modern 700W to 1000W processors. Liquid coolants, by contrast, possess a heat transfer efficacy up to 3,000 times greater than air, allowing for a radical reduction in the infrastructure required to manage equivalent thermal loads. As rack densities push toward 100 kW and projected 600 kW loads by 2027, the CDU becomes the essential mechanism for decoupling the sensitive IT environment from the facility-level cooling infrastructure.

The primary role of a CDU is to manage two distinct cooling loops: the Facility Water System (FWS) and the Technology Cooling System (TCS). The FWS is typically a building-wide circuit that transports heat to external chillers, cooling towers, or dry coolers. The TCS is a closed-loop system that circulates highly controlled, treated coolant—often a mixture of deionized water and propylene glycol (PG25)—directly to the heat sources within the IT racks, such as cold plates or immersion tanks. By maintaining this separation, the CDU protects the high-value IT equipment from the potential contaminants, mineral deposits, and corrosive elements found in facility-level water.

 

 

 

Architectural Foundations: In-Rack vs. In-Row Configuration
The differentiation between in-rack and in-row CDUs is rooted in the deployment strategy and the scale of the compute cluster they are intended to support. An in-rack CDU is a localized solution, typically integrated into the server cabinet itself, whereas an in-row CDU is a standalone enclosure positioned within a row of racks to serve multiple units.

In-Rack CDUs: Precision and Localization

In-rack CDUs are engineered for high-density environments where floor space is at a premium or where liquid cooling must be deployed incrementally. These units are often designed with a standard 4U to 10U rack-mount form factor, allowing them to be installed at the base, top, or side of a server rack. This proximity to the IT equipment minimizes the length of the secondary fluid loop, which significantly reduces the head pressure requirements of the pumps and lowers the risk of pressure drops throughout the system.

 

The localized nature of the in-rack CDU creates an isolated failure domain. In this architecture, a mechanical failure or a leak within a single CDU only impacts the rack it serves, rather than compromising an entire row of servers. This makes in-rack solutions particularly attractive for edge computing sites, laboratory environments, and legacy data centers where a limited number of high-density racks must coexist with traditional air-cooled equipment. Furthermore, in-rack systems are often "plug-and-play," simplifying the rollout of liquid-cooled infrastructure without requiring extensive modifications to the facility’s centralized plumbing.

 

 

In-Row CDUs: Centralized Scalability

In-row CDUs represent the industrial-scale approach to thermal management. Positioned as standalone cabinets between IT racks or at the end of a row, these units are designed to support massive thermal loads, often ranging from 1 MW to 2.5 MW. A single in-row CDU can provide coolant distribution for up to 10 high-power racks, serving as a centralized hub for a large compute pod or an entire data center hall.

 

The primary advantage of the in-row architecture is its efficiency at scale. By centralizing the cooling infrastructure, operators can reduce the total number of CDUs required, which simplifies the overall facility monitoring and maintenance regime. In-row units do not occupy valuable vertical "U-space" within the server racks, allowing for higher compute density per cabinet. However, this centralization introduces a broader failure domain; a malfunction in an in-row unit could potentially affect several racks simultaneously, necessitating a more robust redundancy strategy at the unit level.

Specification	In-Rack CDU (L2L)	In-Row CDU (L2L)
Typical Cooling Capacity	15 kW – 300 kW	550 kW – 2.5 MW
Form Factor	4U – 10U Rackmount	Standalone Cabinet
Typical Deployment	Single Rack / Edge	Cluster / Row / Hyperscale
Failure Domain	Isolated to one rack	Multiple racks in a row
Floor Space Impact	Low (Internal to rack)	High (Requires row space)
Fluid Loop Management	Shortest possible paths	Extensive manifold networks

 

Technical Analysis of CDU Performance Metrics

The performance of any CDU, regardless of its form factor, is governed by several critical technical parameters: thermal capacity, flow rate, approach temperature, and pressure management. These factors determine how effectively the unit can transport heat away from the silicon and into the facility water system.

 

Thermal Capacity and Heat Transfer

Thermal capacity, measured in kilowatts (kW) or megawatts (MW), is the most fundamental metric for CDU selection. Engineers must size the CDU to handle the maximum Thermal Design Power (TDP) of all components in the loop, including CPUs, GPUs, and high-speed networking switches. For AI and HPC clusters, the heat load is often intensely concentrated, with modern GPUs requiring flow rates of 1.5 to 2 liters per minute per kilowatt of heat.

 

The heat transfer process within the CDU is typically facilitated by a high-efficiency plate heat exchanger (PHE). The efficacy of this exchange is measured by the "approach temperature," which is the difference between the temperature of the facility water entering the CDU and the temperature of the coolant being supplied to the IT equipment. A narrower approach temperature—such as 3°C to 5°C—indicates a more efficient heat exchanger, allowing the system to operate with warmer facility water. This capability is crucial for implementing "free cooling" strategies, where the facility can bypass mechanical chillers and rely solely on dry coolers or cooling towers to reject heat to the ambient environment, dramatically reducing the Power Usage Effectiveness (PUE).

 

Flow Rate and Pressure ManagementPumping technology is the operational heart of the CDU. To move coolant through the intricate micro-channels of cold plates and the various bends of a manifold system, the CDU must generate sufficient "head pressure". In-row CDUs, such as the Eaton Boyd ROL4000, are designed to provide up to 80 PSI of pressure to drive flow through high-performance cold plate loops and control valves across multiple racks.Precise flow control is essential for maintaining stable operating temperatures.

Modern CDUs utilize Variable Frequency Drives (VFDs) to adjust pump speeds in real-time based on the compute load. As servers transition from idle to peak processing, the CDU increases the flow rate to manage the sudden thermal spike, ensuring that the silicon stays within its specified "case temperature". This dynamic response is especially critical for AI workloads, which can surge from 4 MW to 130 MW in a matter of milliseconds.

 

Model	Application	Capacity	Pump Redundancy	Key Technical Metric
Supermicro LCS-SCDU-250	In-Rack	250 kW	2N / N+1	Supports 45°C warm water
Supermicro LCS-SCDU-1K3	In-Row	1.8 MW	N+1	1200 LPM Flow Rate
DCX HYDRO CDU1	In-Rack	150 kW	Dual Redundant	60°C Return Temperature
DeepCoolAI DCAI-IWL	In-Row	2.5 MW	N+1	341 GPM Capacity
Eaton Boyd ROL4000	In-Row	2.0 MW	N+1	80 PSI Head Pressure
LiquidStack CDU-1MW	Gallery/Row	1.35 MW	N+1	25% TCO improvement
ACT In-Rack	In-Rack	15 kW	Variable	Two-Phase Dielectric

 

 

Liquid-to-Liquid vs. Liquid-to-Air Rejection Mechanisms

The deployment environment often dictates whether a liquid-to-liquid (L2L) or liquid-to-air (L2A) CDU is most appropriate. These two types represent fundamentally different approaches to final heat rejection.

Liquid-to-Liquid (L2L) Dynamics
L2L CDUs are the gold standard for high-density cooling efficiency. They require a direct connection to a facility's chilled water system, allowing heat to be transferred from the IT coolant to the facility water through a heat exchanger. This architecture provides the highest cooling performance and is capable of supporting the most power-dense racks in hyperscale and AI environments. L2L CDUs are integral to reaching ultra-low PUE, as they facilitate the use of warm water for cooling, which can be reused for district heating or other industrial processes.

Liquid-to-Air (L2A) Flexibility
L2A CDUs, often referred to as "sidecars" when in-rack or standalone units when in-row, disperse heat through a heat exchanger coil and fan system. This configuration is designed for facilities that lack the liquid plumbing infrastructure to support L2L systems. While L2A CDUs simplify installation and reduce initial costs by leveraging existing air-cooling infrastructure, they are inherently limited by the thermal capacity of the facility's air. They effectively transfer the thermal load from the server chips into the room's air, which must then be managed by traditional Computer Room Air Conditioning (CRAC) units. L2A systems are ideal for rapid rollouts, mixed-environment data halls, or edge sites where liquid-to-chip cooling is required but facility-wide plumbing is not feasible.

 

 

Redundancy and Mission-Critical Reliability

In a data center environment where downtime is measured in thousands of dollars per minute, the reliability of the CDU is paramount. Both in-rack and in-row units incorporate multiple layers of redundancy to ensure continuous operation.

 

Component-Level Redundancy
High-end CDUs feature redundancy for all critical moving parts and electronic controllers. This typically includes:
Pumps: Most enterprise CDUs employ N+1 or 2N redundant pump architectures. In an N+1 system, an extra pump remains on standby or operates at reduced speed, ready to take over full load if another pump fails.

Power Supplies (PSUs): Dual power feeds with redundant PSUs ensure that a failure in one electrical circuit does not interrupt the cooling system.

Micro-Control Units (MCUs): Redundant controllers prevent a software or logic failure from shutting down the thermal management system.

 

Operational Continuity and Maintenance

The design of the CDU must allow for "concurrent maintainability," meaning that technicians can service the unit while it is under full load. For instance, in-row units like the Supermicro 1.8 MW CDU integrate an Automatic Transfer Switch (ATS) for power and hot-swappable pump components, ensuring that power and coolant flow remain functional during maintenance. Similarly, filtration systems are often designed with dual-filter bypasses, allowing for the cleaning or replacement of filters without stopping the secondary loop's circulation.

 

 

Risk Mitigation: Leak Detection and Fluid Integrity

The presence of liquid in proximity to high-value IT equipment introduces a unique set of risks, primarily related to leaks and fluid contamination. CDUs mitigate these risks through advanced sensing technology and rigorous filtration.

 

Advanced Leak Detection Strategies
A robust leak detection system is a multi-layered defense mechanism.

 

Sensing Cables: Conductive sensing cables are typically run along the floor beneath the server racks, manifold connections, and the CDU itself. When liquid bridges the cable's conductive threads, the system pinpoint's the leak's location and triggers an alarm.

Spot Sensors: Point sensors or "pucks" are placed at high-risk locations, such as directly under quick-disconnect fittings, valves, or condensate pumps, where even a small drip must be detected immediately.

Mechanical Monitoring: The CDU's internal logic monitors pressure drops and flow rate anomalies. A sudden decrease in pressure in the TCS loop can indicate a major rupture, allowing the system to automatically close isolation valves and stop the pump, limiting the total volume of coolant released.

 

Fluid Quality and Filtration

The integrity of the coolant loop is vital for the longevity of the server cold plates. CDUs are equipped with high-efficiency filters, typically 50 microns, to keep the secondary loop free of debris and contaminants. Operators must carefully balance filtration levels; while finer filters (e.g., <25 microns) provide higher purity, they can increase pressure drop and potentially strip necessary chemical additives from the coolant. Maintaining the proper chemical balance, such as the concentration of propylene glycol and corrosion inhibitors, is essential to prevent biological growth and galvanic corrosion within the micro-channels of the liquid cold plates.

 

 

Operational Integration and Mechanical Management

Beyond thermal performance, the successful deployment of in-row or in-rack CDUs depends on the physical management of the rack environment, specifically cable and hose routing.

 

Operational Integration and Mechanical Management

Beyond thermal performance, the successful deployment of in-row or in-rack CDUs depends on the physical management of the rack environment, specifically cable and hose routing.

Hose Management and Connectivity

The introduction of liquid cooling manifolds and hoses creates a significantly more congested environment within the rack. Proper hose management is critical to avoid obstructing airflow to non-liquid-cooled components and to prevent mechanical stress on connectivity points.

 

Universal Quick Disconnects (UQD): The industry has moved toward standardized UQD interfaces for the secondary loop connections between the manifold and the server. These connectors are designed for thousands of make-and-break cycles with minimal fluid loss, ensuring that server maintenance remains safe and efficient.

Bend Radius and Tensile Loads: Hoses must be routed to maintain their specified bend radius; excessive bending can cause material fatigue or restrict coolant flow. Clamping strategies must account for "hose creep," where materials relax over time due to thermal cycling, potentially leading to leaks at barb interfaces if not properly secured.

 

Cable Management Best Practices

Efficient cable management supports thermal efficiency by ensuring that power and data lines do not block the exhaust of any air-cooled components. In a liquid-cooled rack, this is particularly challenging as hoses and cables compete for space in the vertical patching areas. Best practices include:

Separation: Power and data cables should be isolated on opposite sides of the rack to reduce electromagnetic interference (EMI) and signal crosstalk

Labeling and Color-Coding: Clear labeling and the use of red/blue color-coding for supply and return lines are essential for reducing technician error during rapid troubleshooting.

Strain Relief: Securing cables and hoses to the rack structure removes tension from sensitive connectors and transceivers, preventing accidental disconnections.

 

 

Economic Analysis: Total Cost of Ownership and ROI

While liquid cooling architectures generally require higher initial capital investment (Capex) than traditional air-cooled systems, their long-term operating expenditure (Opex) provides a compelling return on investment (ROI).

Capex vs. Opex Dynamics
Air-cooled facilities are simpler and cheaper to build upfront but become exponentially more expensive to operate as densities increase beyond 20 kW per rack. Liquid cooling systems, whether in-rack or in-row, significantly lower energy costs over time by:

Reducing Fan Power: In-row cooling can save more than 50% in fan power consumption compared to room-based cooling.

Optimizing Chiller Usage: By supporting higher facility water temperatures (W3/W4 standards), liquid cooling enables the use of efficient dry coolers for a larger portion of the year.

Maximizing Floor Space: In-rack CDUs allow for greater compute density within the same physical footprint, potentially deferring the need for expensive facility expansions.

 

Metric	Air Cooling	In-Row Liquid Cooling	In-Rack Liquid Cooling
Upfront Cost (Capex)	Lowest	Moderate-High	Highest per unit
Energy Cost (Opex)	Highest at density	Low	Lowest at extreme density
Fan Energy Savings	Baseline	>50% Reduction	>50% Reduction
Maintenance Overhead	Low	Moderate (Centralized)	High (Distributed)
Scaling Flexibility	Low	High (Cluster-level)	Highest (Rack-level)

 

 

 

Strategic Decision Matrix: Selecting the Optimal Architecture

The decision between in-row and in-rack CDUs is not a matter of one being inherently superior; rather, it is a matter of alignment with specific operational goals and facility constraints.

 

Use Cases for In-Rack CDUs：

In-rack CDUs are the preferred choice for:

Edge Computing and Remote Nodes: Where space is limited and cooling must be self-contained.Heterogeneous

Data Halls: Where liquid-cooled AI pods must exist alongside traditional air-cooled racks.

Modular Growth: For organizations that prefer to scale their investment rack-by-rack as compute demands increase.

Retrofit Scenarios: Where adding centralized row-level plumbing is physically impossible or cost-prohibitive.

 

Use Cases for In-Row CDUs：

In-row CDUs are the optimal choice for:

Hyperscale and Large-Scale AI Clusters: Where thousands of GPUs are deployed in consistent, high-density blocks.

Greenfield Data Centers: Where liquid-ready infrastructure can be designed from the ground up to support centralized distribution.

High-Efficiency Operations: Where minimizing the total number of managed units and maximizing economies of scale is a primary goal.

Extreme Density (>300 kW per cluster): Where individual in-rack units would be insufficient or would consume too much internal rack space.

 

 

 

Technological Frontiers: Two-Phase Cooling and Beyond

As the industry looks toward 2027 and the reality of 600 kW racks, the next evolution of the CDU is already emerging: two-phase liquid cooling.

Mechanics of Two-Phase Systems
Unlike traditional single-phase systems where the coolant remains liquid, two-phase cooling utilizes specialized dielectric fluids that boil and vaporize at the heat source. This phase change absorbs significantly more energy than simple temperature increase. The resulting vapor travels to the CDU, where it is condensed back into a liquid and recirculated. Two-phase CDUs, such as those from ACT, offer several advantages：

Lower Pumping Power: The latent heat of vaporization allows for much lower mass flow rates compared to single-phase water-glycol systems.

Uniform Thermal Performance: The phase change occurs at a constant temperature, ensuring that all chips in a dense cluster remain at the exact same temperature.

Enhanced Safety: The use of dielectric fluids ensures that even in the event of a leak, there is no risk of electrical short circuits

 

 

Key Points:

Lian Li manufactures In-Row CDUS and In-Rack CDUs, offering both Air-to-Liquid and Liquid-to-Liquid cooling solutions. These units are compatible with servers from NVIDIA, AMD, and Intel—including NVIDIA's GB200/GB300 and B200/B300 series—and all products are certified under RoHS, CE, and UL standards.

 

 

Conclusion: Orchestrating the Thermal Future

The selection between in-row and in-rack CDUs is a foundational decision that shapes the resilience, efficiency, and scalability of the modern data center. In-rack architectures provide the precision and localized risk management necessary for edge sites, modular expansions, and diverse IT environments. They offer an unparalleled "pay-as-you-grow" flexibility that appeals to organizations navigating the early stages of the liquid cooling transition.

 

In-row architectures, conversely, provide the raw capacity and centralized management efficiency required to support the massive AI and HPC clusters that define the current computing era. By leveraging economies of scale and industrial-grade redundancy, in-row CDUs enable hyperscale operators to push the boundaries of compute density while maintaining strict control over PUE and operational complexity.

 

As thermal loads continue their inexorable climb, the CDU will remain the vital interface between the digital and physical worlds. Architects must look beyond immediate cooling needs and evaluate these systems through the lens of total cost of ownership, risk mitigation, and future technological compatibility. Whether deploying a single rack at the edge or a multi-megawatt cluster in a centralized hall, the mastery of CDU architecture is the key to unlocking the full potential of high-performance computing.

 

 

FAQ :

What is the main difference between an in-row CDU vs. in-rack CDU?
The primary difference lies in their installation footprint and structural capacity. An in-row CDU is a floor-mounted, standalone unit designed to sit directly within a row of server racks, providing centralized liquid cooling loops for multiple high-density racks simultaneously. Conversely, an in-rack CDU is a compact, modular enclosure integrated directly inside a single 19-inch server rack, dedicating its total cooling distribution exclusively to that specific rack's components .

 

When should data center operators choose an in-rack CDU over row-based alternatives?
An in-rack CDU is the ideal choice when retrofitting legacy air-cooled facilities or selectively deploying isolated high-performance computing (HPC) and artificial intelligence (AI) clusters . Because it resides completely within the server enclosure, it does not require dedicated white space on the data center floor. It allows facility managers to pilot direct-to-chip liquid cooling architectures on a per-rack basis without overhauling the broader row layout or facility-level fluid networks.

 

How do CDUs improve overall data center cooling efficiency for AI workloads?
Modern data center cooling strategies rely heavily on Coolant Distribution Units (CDUs) to isolate secondary technology cooling loops from primary facility water systems, protecting sensitive electronics while managing extreme heat . As AI hardware pushes chip power densities beyond the limitations of traditional air-conditioning, CDUs deliver precise flow, pressure, and temperature control directly to cold plates or rear-door heat exchangers . This targeted liquid delivery dramatically reduces reliance on energy-heavy chillers, slashing the facility's Power Usage Effectiveness (PUE) .