Buyer's Guide_
How to Choose a GPU Deployment Partner: A Buyer's Checklist
This checklist specifies the rack-assembly, cabling, cooling, and commissioning capabilities plus proof points that data-center operators must verify before engaging a physical-layer GPU deployment crew for NVL72-class builds.
Key facts
- NVLink in NVL72-class racks runs exclusively over internal copper spines or backplanes; scale-out InfiniBand or Ethernet uses separate MPO/fiber trunks between switches and racks.
- MPO trunk cables arrive factory-terminated and polished; field crews perform only routing, cleaning, inspection with a fiber microscope, and end-to-end continuity or OTDR testing.
- Liquid-cooling loops require pressure testing to the OEM-specified value, followed by flow balancing and leak checks before GPU power-on.
- Structured cabling must meet separation rules between power and signal pathways per the relevant TIA or ISO standards for data-center pathways.
- Commissioning includes full rack power sequencing, GPU-to-switch link training, and thermal validation under load before customer acceptance.
- A single mis-routed or contaminated MPO connector can drop an entire leaf-spine fabric; inspection before every insertion prevents this.
- Reference projects must include at least one completed rack-scale deployment with documented test results for both copper NVLink and fiber scale-out networks.
Rack Assembly Experience with NVL72-class Systems
Demand evidence that the crew has performed full mechanical integration of NVL72 racks, including placement of the NVLink copper spine, power shelf alignment, and manifold connections for liquid cooling. Ask for the exact sequence they follow when seating the compute trays and torquing the bus-bar connections; the correct order prevents mechanical stress on the backplane before any power is applied.
The partner must also demonstrate they can read and execute the OEM rack-level bill of materials without substitution. Any deviation in rail depth, PDU orientation, or manifold routing must be flagged and approved by the operator before proceeding, because downstream cooling and cable routing depend on these dimensions.
Structured Cabling and MPO Handling Practices
Confirm the crew routes MPO trunks in dedicated pathways separate from power cables and maintains minimum bend radii at every turn. They must clean and inspect every connector face with a calibrated microscope before mating; skipping this step is the most common source of intermittent link flaps later.
Require written proof that they perform both continuity and insertion-loss testing on every fiber strand after installation. The test records become part of the handover package so operators can baseline future changes.
Liquid Cooling Integration and Commissioning
The crew must isolate each rack manifold, perform a hydrostatic pressure test to the OEM limit, and hold that pressure while monitoring for drops. Only after passing this test do they introduce coolant and balance flow rates across the GPUs using the rack-level flow meters.
They must also verify that quick-disconnect fittings are fully seated and that air is purged from the loop before any GPU power is applied. Residual air pockets cause localized overheating that only appears under full load.
Network Validation and Testing Protocols
After physical installation, the partner runs link-training sequences on the scale-out fabric and confirms every port reaches the expected speed and FEC state. They must also execute a full rack-to-rack ping and bandwidth test using the operator’s chosen traffic generator before declaring the fabric ready.
Because NVLink remains internal to the rack, the validation focuses on the fiber domain only; any claim that fiber affects NVLink status indicates the crew does not understand the topology.
Common Failure Modes in GPU Deployments
The most frequent field failures are contaminated MPO end-faces and reversed fiber polarity. Both produce high bit-error rates that only surface after the cluster is placed under training load, forcing costly re-work once racks are already populated.
Another recurring issue is improper manifold torque or missed O-ring seating during liquid-cooling installation; these leaks appear after thermal cycling and can damage adjacent electronics. Crews that skip the post-install pressure re-test or fail to document every connector inspection are the primary source of these problems. Operators should require the partner to show their inspection logs and pressure-test certificates before any rack is energized.
Reference Checks and Documentation Requirements
Request contact information for at least two completed projects of similar scale and ask the prior operator whether the crew delivered test results in a usable format and responded to punch-list items within the agreed window. Leviathan Systems maintains this level of documentation on every engagement.
Insist on a sample handover package that includes rack elevation drawings, fiber test reports, cooling loop certificates, and as-built cable schedules. Absence of any of these items indicates the partner treats commissioning as an afterthought rather than a required deliverable.
Standards referenced: TIA-942 data-center cabling pathways · IEC 61754-7 MPO connector interface · OEM-specified hydrostatic test pressure for liquid manifolds
Frequently asked_
How do I confirm a crew understands that NVLink stays inside the rack?
Ask them to describe the copper spine routing and the separate fiber domain used for the leaf-spine fabric. Any answer that mixes the two domains shows they have not worked on NVL72-class systems. Request a marked-up rack elevation that labels both domains correctly.
What test equipment must the partner bring for MPO trunks?
A calibrated fiber microscope for end-face inspection, an MPO continuity tester or OLTS, and an OTDR with the correct launch cable. They must produce the calibration certificates for each instrument before work begins.
How long should the liquid-cooling pressure test hold?
Follow the exact duration and pressure value listed in the OEM rack installation manual. Record the starting and ending pressure plus ambient temperature; any measurable drop requires the loop to be drained and re-sealed before proceeding.
What documentation should appear in the final handover?
Rack elevations, fiber test reports with pass/fail values, cooling pressure and flow certificates, and a signed checklist confirming every GPU tray was seated and powered in sequence. Leviathan Systems includes these items as standard deliverables.