Data Center Design_
White-Space Planning for Liquid-Cooled GPU Halls
Plan rack rows, CDU interfaces, service access and IT cable pathways around the selected GPU system before accepting a white-space layout.
Key facts
- Row pitch uses a defined rack-grid reference. Manifold projections, doors, hoses and equipment-removal space are separate layout inputs.
- In NVIDIA's DGX GB rack guide, NVLink uses an internal passive copper cable-cartridge backplane. External optical links can connect switches to adapters as well as other switches. [N1, N2, N3]
- CDU location, support, connection exits and service space depend on the selected configuration; one CDU per two rows is not a general planning rule. [V1]
- A liquid-to-liquid CDU separates facility and technology cooling circuits. Row orientation must account for both routes and the equipment's remaining airflow needs. [A1, N1]
- Coupling access and leak-protection provisions belong to the applicable equipment and site design; neither universal drip-pan placement nor a generic valve interval is established here.
- End-face inspection, polarity verification and optical-loss measurement provide different evidence. An inspection result alone does not establish a complete optical channel's performance. [F1]
Row Spacing and Rack Orientation
Row spacing starts with the selected rack dimensions and the front and rear access requirements. Define row pitch between consistent rack reference lines, such as the centerlines shown on the project plan. Then show the occupied outline, including external manifolds, rear doors and other projections, separately from the space needed to open, service or remove equipment. An internal cold plate does not by itself establish extra external rack depth.
Orientation must address air and liquid paths together. NVIDIA's DGX GB hardware guide, updated September 17, 2026, describes CPU and GPU cold plates while networking and storage components in the compute tray remain air cooled. A liquid-cooled description therefore does not eliminate the airflow question. Use the actual equipment's orientation and environmental requirements rather than orienting every cold aisle toward a facility-water entry. [N1]
For a liquid-to-liquid CDU, trace the facility connection to the CDU separately from the technology-circuit distribution to the racks. ASHRAE's 2019 white paper explains this circuit separation. A shorter facility route does not establish a shorter rack-side route or a lower pressure drop for the complete arrangement. Compare the proposed orientation against the selected connection locations, service access and the cooling designer's applicable routing basis. [A1]
CDU and Manifold Placement
End-of-row placement and an adjacent mechanical bay are layout candidates. Their suitability depends on the actual CDU configuration, both liquid interfaces, support requirements and access. Obtain the applicable equipment outline and service drawings before reserving a position; a footprint alone leaves connection and maintenance space unresolved.
Vertiv's publicly available 2022 XDU450 Application and Planning Guide illustrates the configuration problem. It identifies wheels and adjustable feet, top or bottom connection exits, and optional six-way or eight-way manifolds. These are options for that named product, not a universal housekeeping-pad requirement, three-sided clearance or rack count. Confirm the offered variant and current applicable OEM information for the project. [V1]
The cooling designer must establish the served equipment, thermal and hydraulic basis, distribution and isolation boundaries. Valve spacing alone cannot establish that an individual rack can be removed from service independently. Show the specialist-supplied interface locations and access needs in the physical IT plan, including any dependency on rack positioning. Coupling type, fluid compatibility, leak protection and the actual installation sequence remain part of the selected cooling package.
Service Clearances and Access Paths
Protect the selected OEM's applicable service space from fixed equipment, trays and temporary staging. Where the product has a removable heat exchanger, panel or other service assembly, show its required removal path as well as the closed cabinet. At the rack, consider the relevant door position, equipment extraction and permitted hose movement. A hose loop or door that fits in the operating view can still obstruct the service view.
Compare those spaces with IT cable access, electrical working space, egress and equipment delivery routes. Include height where an overhead tray or pipe could intersect the required movement. A cooling route has no automatic priority over these other requirements. Record each conflict for the appointed equipment, facility and design owners to resolve against the applicable OEM documents and project requirements.
Clear access does not establish that maintenance can occur without interrupting rack operation. That depends on the actual cooling arrangement, permitted equipment state and responsible parties' service provisions. Keep the spatial question and the operational question separate when accepting the layout.
Scale-Out Network Integration
Include the actual host and switch endpoints in the room cable plan. NVIDIA's DGX GB networking guide, updated September 17, 2026, distinguishes its within-rack NVLink fabric from compute, storage, provisioning and management networks. Its hardware guide identifies the internal NVLink passive copper cable-cartridge backplane. Those product-specific descriptions do not mean that all GPU systems keep NVLink within one rack or that only external scale-out links affect platform operation. [N1, N2]
NVIDIA's LinkX 100G-PAM4 overview, updated November 18, 2025, describes optical switch-to-adapter and switch-to-switch arrangements. MPO trunks therefore cannot be limited to connections between switches. Select the actual transceivers, connector type and cable assembly from the approved network design; the MPO label alone does not establish compatibility. [N3]
Coordinate cable pathways and patch-panel access with the supplied cooling service spaces. Preserve access to both domains without imposing a universal installation order or requiring patching on the CDU side. Changes to a tray, panel or endpoint can affect service access even when the rack positions remain fixed.
Keep optical evidence distinct too. Fluke's MultiFiber Pro page separates loss and polarity measurement from FiberInspector end-face inspection. An inspection image or continuity result cannot supply a missing loss measurement. The network owner defines the required channel evidence and its acceptance basis; cooling readiness and network readiness retain their own prerequisites in the project schedule. [F1]
Common Failure Modes in White-Space Layouts
A service face placed against a wall, or a tray added inside a removal path, can make a nominally spacious layout unusable for the intended task. Compare every relevant fixed object with the selected equipment's service drawing, then retain the affected view and disposition. Do not infer the consequence or a shutdown requirement from the obstruction alone.
Consider a fictional layout in which the selected OEM drawing requires 1,200 mm of clear service depth from the rear cabinet face over a defined width and height. The plan provides exactly 1,200 mm from that face to the opposite fixed boundary. A proposed cable-management assembly projects 200 mm into that same space across the required envelope. The remaining clear depth is 1,000 mm, leaving a 200 mm conflict against the stipulated requirement. These invented dimensions are not a recommended aisle width or a product specification.
The rack-grid pitch has not changed, but the service space has. Removing the projection, changing its position or revising the equipment arrangement are proposals for the responsible owners to assess, not automatic remedies. Merely moving a centerline label does not resolve the geometry. This is why an occupied outline and a service-state view are useful alongside the rack grid.
Unresolved hose geometry or leak exposure near IT equipment also belongs in the interface review. A plan showing a low point does not establish a pump-trip diagnosis, and a drip pan shown under one component does not establish complete leak protection. Obtain the cooling owner's applicable layout and protection provisions instead of adding generic vent or pan instructions.
Pre-Commissioning Layout Handoff
Before accepting the layout for its intended use, reconcile the equipment identities and revisions, rack references, occupied outlines, specialist-supplied service spaces and IT pathways. Preserve marked-up drawings and any agreed field measurements with their locations and reference points. Distinguish a planned dimension from an observed dimension and from the responsible owner's acceptance decision. No site observation or acceptance result is represented by this article.
Identify which prerequisites and records belong to the cooling, facility, OEM and IT owners. ASHRAE's 2019 white paper distinguishes manufacturer-supplied assemblies from site-installed interconnecting piping or hoses. Its historical discussion explains why their evidence boundaries differ; it is not a current test method. A factory record does not automatically cover later site joints. Actual methods, applicable codes, test scope and permission to proceed must be established by the appointed qualified owners. [A1]
Leviathan's planning offer concerns agreed physical IT layouts, cable routing and documentation, with authoring capability, formats and reviewers confirmed per engagement. It excludes licensed or regulated facility engineering and all liquid-side design, work, coordination, subcontracting, measurements and commissioning. The customer supplies the cooling and facility requirements and their responsible owners' dispositions for the physical IT package.
Use the linked Physical Infrastructure Planning & CAD service and project inquiry to discuss that package. Share the selected equipment list, current plans, approved specialist interface information, intended document use, requested outputs and unresolved access conflicts so scope, requirements and availability can be assessed.
Frequently asked_
How far apart should CDUs be placed when one unit serves two rows?
There is no general spacing or one-unit-per-two-rows rule. The appointed cooling designer needs the selected CDU's applicable thermal and hydraulic basis, served equipment, distribution and isolation arrangement, routes and service access. Hose length is one input; it cannot determine the quantity or position by itself.
Do liquid manifolds share the same pathway as MPO trunks?
The coordinated design must show their supports, service access and any shared or crossing space. Separate routing can preserve access to each domain, but a universal separation distance, vertical order or installation sequence does not follow from this article. Resolve the actual arrangement with the appointed cooling, network and facility owners.
What clearance is required behind a rear-door heat-exchanger rack?
Use the applicable cabinet, door and hose configuration's OEM information, including the relevant open-door and service or removal states. Compare those requirements with rear equipment, cable exits and site access. The closed-door footprint alone does not establish sufficient rear clearance.
When is the static pressure test performed relative to rack installation?
That sequence must come from the applicable approved cooling package and its responsible specialists. Supplied assemblies and site interconnections can have different evidence boundaries. This layout article supplies no pressure value, coupling-cycle count, pump setting or universal test sequence. [A1]
Who signs off on final white-space clearances?
The customer must identify the appointed acceptance owners for the actual equipment, facility requirements and physical IT package. Retain the applicable drawing revision, reference points, recorded observations, open conflicts and responsible disposition. A marked-up plan or photograph does not itself confer cooling, electrical or overall site approval.