Buyer's Guide_
RFP & SOW Checklist for a GPU Build: Scope It So Nothing Falls Through
This RFP and SOW checklist specifies the exact line items needed to scope GPU rack deployments from rigging through acceptance for H100 to GB300 NVL72 systems, ensuring no gaps in mechanical, cabling, cooling, or verification work.
Key facts
- NVLink GPU-to-GPU connectivity in NVL72-class racks occurs exclusively over internal copper backplanes and spines, separate from any fiber infrastructure.
- MPO trunk cables for scale-out InfiniBand or Ethernet networks arrive factory-terminated; field work consists of routing, cleaning, inspection, and testing only.
- Fiber end-face inspection must follow IEC 61300-3-35 criteria before any MPO connection is made.
- Liquid cooling loops require sequential pressure testing, flow verification, and leak checks per OEM procedures prior to rack power-up.
- Rack power and grounding integration must align with the data hall's electrical single-line drawings and grounding electrode system.
- Structured cabling pathways follow TIA-942 space and pathway guidelines for separation of power and data runs.
- Acceptance requires documented results from calibrated test equipment including OTDR traces and continuity testers matched to the installed media.
Rack Rigging and Mechanical Integration
Rigging begins with verification of floor loading against the data hall structural drawings and confirmation that the rack base plates or casters match the intended final position. Crews then lift and set each rack using rated equipment, aligning mounting holes to the floor anchors before torquing to the values listed in the rack OEM installation manual.
After placement, installers check plumb and level across the rack frame and adjacent units to maintain aisle containment integrity. Any deviation requires shimming before proceeding to internal component installation because misalignment transmits stress into the liquid cooling manifolds and backplane connections.
Structured Cabling for Scale-Out Networks
Scale-out cabling uses MPO-terminated trunks routed through overhead or underfloor pathways sized per TIA-942. Installers pull trunks with adequate slack at both ends, maintain minimum bend radii during routing, and secure cables to ladder rack or basket tray without over-compression.
Once positioned, each MPO connector undergoes cleaning with the OEM-approved cassette and inspection under a calibrated microscope meeting IEC 61300-3-35 pass/fail limits. Only connectors that pass are mated to patch panels or switch ports; failed ends are set aside for replacement rather than forced into service.
Liquid Cooling Loop Commissioning
Cooling commissioning starts with visual inspection of all manifold connections and quick-disconnects for proper seating. Technicians then perform a low-pressure nitrogen hold test followed by a water or coolant fill under controlled flow, monitoring for pressure decay over the duration specified in the cooling distribution unit manual.
Flow meters and temperature sensors are verified against the building management system before any GPU nodes receive power. Leaks discovered at this stage are isolated and repaired; the loop is retested until stable because even minor drips will corrode electrical components once the system is energized.
GPU Networking and Internal Connectivity Verification
Internal NVLink verification occurs after all nodes are seated and powered; crews run the OEM diagnostic suite to confirm link status across the copper backplane. No fiber or MPO connections participate in this step because NVLink remains entirely intra-rack.
Scale-out ports are tested separately with an OTDR or MPO continuity tester to produce baseline loss and length records. Results are compared against the link budget calculated from the trunk specifications; any port exceeding the budget triggers re-inspection of the mated pair.
Common Failure Modes in Field Deployment
The most frequent failure is contamination on MPO end faces introduced during pulling or patching; dust trapped between ferrules produces intermittent high loss that only appears under load. Catching this requires 100 percent inspection with a microscope before every mating rather than relying on a single pre-pull check.
Another recurring issue is mismatched torque on rack mounting hardware or manifold clamps, leading to vibration-induced leaks or backplane misalignment after the system is online. Field crews prevent this by using calibrated torque wrenches and documenting each fastener against the OEM table before moving to the next rack.
Improper separation of power and data pathways during installation creates EMI coupling that surfaces only during full-system burn-in. The remedy is strict adherence to the TIA-942 separation distances marked on the pathway drawings and verification with a final walkthrough before acceptance.
Power, Environmental, and Acceptance Documentation
Power-up sequencing follows the rack OEM checklist, confirming each PDU feed, branch circuit, and ground bond before applying load. Environmental sensors are cross-checked against ASHRAE thermal guidelines for the target inlet temperature range.
Crews compile all test results, as-built drawings, and serial-number inventories into a single handover package. The customer signs only after every test point meets the acceptance criteria listed in the SOW; open items are tracked to closure before the rack is released for production workloads.
Standards referenced: TIA-942 · IEC 61300-3-35 · ASHRAE thermal guidelines for data centers
Frequently asked_
Does the SOW need separate line items for NVLink testing versus scale-out fiber testing?
Yes. NVLink verification uses the copper backplane diagnostics inside the rack and must be listed as an intra-rack step. Scale-out testing covers MPO trunks and switch ports with OTDR or continuity testers and must appear as a distinct deliverable because the two domains do not overlap.
What acceptance criteria apply to liquid cooling loops before GPU power-on?
The SOW must require a documented pressure hold test, flow verification at design rates, and zero-leak confirmation using the OEM procedure. These steps occur after mechanical installation but before any electrical energization of the nodes.
How should fiber inspection be scoped to avoid later link failures?
Include a requirement for 100 percent end-face inspection to IEC 61300-3-35 limits on every MPO connector before mating. The SOW should also mandate calibrated microscopes and retention of inspection images as part of the handover package.
Should rigging torque values be called out explicitly in the SOW?
The SOW should reference the rack OEM torque table and require calibrated tools plus documented readings for every critical fastener. This prevents vibration or thermal cycling issues that appear only after commissioning.
What documentation must be delivered at acceptance for a GPU rack deployment?
The package must contain as-built pathway drawings, test results from OTDR and continuity testers, cooling loop pressure and flow logs, NVLink and scale-out port status reports, and serial inventories. Sign-off occurs only after all items meet the stated pass criteria. Integrators such as Leviathan Systems deliver this package as standard practice.