Commissioning_
The As-Built & Handoff Package Every GPU Deployment Should Deliver
This article specifies the exact documentation deliverables for GPU rack commissioning, covering asset inventories, test reports, and as-builts that allow operators to bring clusters online without delays or rework.
Key facts
- MPO trunk cables for scale-out InfiniBand or Ethernet networks are factory-terminated and polished; field work consists of patching, routing, cleaning, inspection with IEC 61300-3-35 criteria, and testing with calibrated MPO continuity testers or OTDRs.
- GPU-to-GPU NVLink connectivity in NVL72-class racks occurs exclusively over the internal copper NVLink spine or backplane and does not depend on any fiber or MPO links.
- TIA-606-C governs labeling of all cabling, racks, and panels; labels must include unique identifiers that match the as-built drawings and asset database exactly.
- Power and cooling validation records must capture measured values against OEM nameplate ratings and liquid cooling loop parameters before any compute nodes are powered.
- Asset lists require serial numbers, MAC addresses, and firmware versions captured at rack integration, not from purchase orders.
- Common handover failures occur when polarity on MPO trunks is not verified end-to-end or when rack elevation drawings omit actual installed positions of switches and manifolds.
- Leviathan Systems delivers these packages as part of every rack assembly and commissioning engagement.
Rack-Level Asset Inventory
Every rack requires a complete asset list that records the serial number, part number, MAC addresses, and installed firmware for each GPU tray, NVLink switch tray, power shelf, and manifold. These values are captured during integration using barcode scanners tied to the deployment database rather than copied from shipping manifests.
The list must also include the physical rack U-position for every device and the corresponding label ID applied under TIA-606-C rules. Operators later use this data to populate DCIM systems and to confirm that replacement parts match the original configuration during maintenance.
Without this baseline, day-one troubleshooting cannot distinguish between a mis-cabled link and an incorrect or swapped component.
Scale-Out Network Test Documentation
All MPO trunk cables carrying the InfiniBand or Ethernet fabric between racks must be documented with insertion loss, return loss, and polarity results obtained from a calibrated MPO tester. Each trunk is tested after cleaning and inspection against IEC 61300-3-35 end-face criteria, and the test file is named with the trunk identifier from the as-built drawings.
Switch port mappings are recorded showing which rack uplink connects to which leaf or spine port. This mapping is verified by lighting the port LEDs or using LLDP before the handoff package is accepted.
NVLink connectivity inside the rack is verified separately through the copper backplane using the OEM diagnostic utilities; fiber test results are never presented as evidence of NVLink status.
Liquid Cooling and Power Validation Records
Cooling loop documentation includes measured flow rates, supply and return temperatures, and pressure differentials at each manifold connection, compared against the OEM-specified operating envelope. Any deviation outside tolerance requires a corrective action entry before the rack is released.
Power shelf commissioning records show input voltage, phase balance, and breaker coordination results plus the actual load drawn by the populated GPUs at staged power levels. These measurements confirm that the rack power distribution matches the design one-line diagram.
Both sets of records are timestamped and signed by the commissioning engineer so that subsequent thermal or electrical events can be correlated to the original baseline.
As-Built Drawings and Rack Elevations
As-built drawings update the original rack elevation to show the exact installed positions of every tray, switch, and manifold, including any field deviations approved during assembly. Cable routing paths for both copper NVLink spines and MPO trunks are shown with actual bend-radius compliance noted at each turn.
Elevation drawings are cross-referenced to the asset list so that a serial number can be located to a precise U-position without opening the rack doors. Separate sheets detail the front and rear cable management for each rack type.
These drawings are delivered in both native CAD and PDF formats with revision control so that future moves, adds, or changes start from verified geometry rather than assumptions.
Common Failure Modes in Documentation Handover
The most frequent failure is mismatched labeling between physical ports and the test reports, which occurs when installers apply labels from a pre-printed sheet that does not reflect last-minute port reassignments. This forces the operations team to retest or trace every link before the cluster can be accepted.
Another common gap is omission of end-to-end polarity verification on MPO trunks; a single reversed pair will pass continuity but will prevent the fabric from training. The error is caught only when the first workload attempts to use those links.
Incomplete firmware version capture leads to later mismatches when nodes are replaced; the handoff package must contain the exact versions observed at commissioning so that drift can be detected during the first health checks.
Operator Handoff and Sign-Off
The final package is reviewed in a joint walk-down where the deployment crew demonstrates that every test report matches the physical installation and that all assets are reachable from the management network. Sign-off occurs only after the operator confirms they can power on the rack, run the OEM diagnostics, and see the expected fabric topology.
Leviathan Systems archives the complete package with the customer and retains a controlled copy for warranty support. Any open items are logged with owner, due date, and verification method before the rack is handed over.
Standards referenced: TIA-606-C · IEC 61300-3-35 · ANSI/TIA-568 series for structured cabling
Frequently asked_
How do we verify NVLink status during handoff?
NVLink is verified exclusively through the internal copper spine using the GPU tray diagnostics and nvidia-smi output. Fiber or MPO test results are never used as proof of NVLink connectivity because the two fabrics are physically separate domains.
What format should MPO test results be delivered in?
Results are delivered as native files from the calibrated MPO tester plus a summary spreadsheet that maps each trunk ID to its measured loss, polarity, and pass/fail status. PDF exports alone are insufficient because operators need to import the data into their own systems.
Who owns updates to the as-built package after initial handover?
The operator takes ownership of the master as-built set. Any subsequent change must be documented with a delta drawing and updated asset entry that references the original commissioning revision so that the baseline remains traceable.
Why is firmware version capture required at rack integration?
Firmware versions determine compatibility with the cluster software stack. Capturing them at integration prevents later disputes about whether a node arrived with the expected revision or was altered during staging.
What happens if a trunk fails polarity testing after installation?
The trunk is flagged, the polarity is corrected by swapping the appropriate MPO connectors at one end only, and the full test suite is repeated. The corrected results replace the original entry in the handoff package with a note explaining the rework.