LEVIATHAN SYSTEMS

Commissioning_

Writing an Acceptance Test Plan (ATP) for a GPU Cluster

Sergey Evstigneev·Field Engineering, Leviathan Systems, GPU rack assembly, structured cabling & commissioning for AI data centers·

This article specifies the exact sequence of sections, pass/fail criteria, and pre-energization checks required in an Acceptance Test Plan for GPU racks, separating internal copper NVLink verification from scale-out fiber work.

Key facts

  • GPU-to-GPU NVLink in NVL72-class racks runs exclusively over the copper backplane inside the rack enclosure.
  • MPO trunk cables for scale-out InfiniBand or Ethernet are factory-terminated; field work consists of routing, cleaning, inspection, and testing only.
  • A calibrated MPO continuity tester or OTDR must be used to verify each fiber strand before any rack receives power.
  • Liquid cooling loops require pressure decay testing and visual inspection of every manifold connection prior to electrical energization.
  • Grounding resistance at the rack level must be verified against the facility grounding electrode system before PDU breakers are closed.
  • All tests in the ATP are executed with breakers open and no GPUs under load.
  • Every test point is documented with photographs and meter readings before proceeding to the next rack.

Scope boundaries and document prerequisites

The ATP must list every rack serial number, the exact GPU model population, and the copper NVLink spine configuration before any measurement begins. It also records the scale-out switch ports and the MPO trunk identifiers that will connect this rack to the rest of the fabric.

Prerequisites include completed rack alignment to the raised floor, verified torque on all busbar connections, and delivery of the facility single-line diagram showing the intended feed path. No test proceeds until these items carry dated sign-off from both the mechanical and electrical leads.

The document states explicitly that internal copper NVLink links are excluded from fiber test requirements and that MPO work addresses only the InfiniBand or Ethernet compute fabric.

Mechanical and structural inspection sequence

Begin with a visual check of every rack rail, shelf, and cable management arm against the OEM rack elevation drawing. Measure and record clearance between the copper NVLink spine and any adjacent liquid cooling manifold to confirm no contact points exist.

Next verify that all blanking panels are installed in unoccupied GPU slots and that rear door perforations align with the cold-aisle containment seals. Any deviation requires correction and re-inspection before the ATP advances.

Document the position of each PDU and the routing of its input whips so that future maintenance can isolate a single rack without disturbing neighboring units.

Power distribution and grounding verification

With all PDU breakers locked open, measure continuity from each rack ground bar to the facility grounding electrode system using a low-resistance ohmmeter. Record the value at the rack ground bar and again at the nearest under-floor ground grid point.

Inspect every power whip for proper torque marks on the compression lugs and confirm phase rotation matches the facility standard using a rotation meter. Any reversal or high-resistance reading stops the ATP until corrected.

The plan requires photographs of each torque wrench setting and the final ground resistance reading before the rack is released for the next stage.

Liquid cooling loop integrity checks

Pressurize each manifold segment to the OEM-specified test pressure and monitor for decay over the required hold period. Walk the entire loop with a leak detector and inspect every quick-disconnect and barb fitting under illumination.

Record inlet and outlet temperatures of the facility chilled water supply at the rack isolation valves to establish baseline delta-T before any heat load is applied. Confirm that drain valves are closed and capped.

Any pressure loss or visible moisture requires isolation of the affected segment, repair, and full retest before the ATP continues.

Copper NVLink and scale-out fiber separation

Confirm continuity across every copper NVLink lane on the internal backplane using the rack management controller diagnostic port while breakers remain open. This test verifies seating of the GPU trays against the spine but does not involve external fiber.

For the scale-out network, inspect each MPO connector end-face with a calibrated microscope, clean when required, and test insertion loss with a calibrated MPO continuity tester or OTDR. Record pass/fail against the link budget defined in the facility design package.

The ATP explicitly states that a failed MPO strand does not affect the internal NVLink copper domain and must be resolved through patching or trunk replacement before rack power-up.

Common failure modes encountered in the field

The most frequent stoppage occurs when an MPO connector is cross-patched between two different fabric planes, producing an OTDR trace that shows the correct length but fails continuity at the far end. This is caught by labeling each MPO leg with both rack and switch port identifiers before the first test.

Another recurring issue is residual pressure in a cooling segment caused by an upstream valve left partially open; the decay test then fails even though no leak exists. The remedy is to isolate the rack valves completely and re-zero the gauge before the timed hold begins.

Ground loops created by accidental contact between the rack frame and an uninsulated busbar joint produce unexpectedly low resistance readings that later cause GFCI trips. These are identified by repeating the ground measurement after each mechanical adjustment.

Sign-off gates before energization

Every section of the ATP must carry dated initials from the field lead and the customer commissioning engineer. No section may be marked complete if any single test point remains outside tolerance. Leviathan Systems personnel archive the completed document with rack serial numbers only after all gates pass.

Once all mechanical, power, cooling, and fiber items are accepted, the next rack is released for the same sequence. Only after the full row or suite passes does the plan authorize closure of the first PDU breaker.

Standards referenced: TIA-568.3-D for fiber optic cabling test methods · IEC 61300-3-35 for connector end-face inspection · ASHRAE TC 9.9 thermal guidelines for data processing environments

Frequently asked_

When is the copper NVLink backplane tested relative to the MPO fiber links?

Copper NVLink continuity is verified first using the rack management controller while all breakers are open. MPO fiber testing for the scale-out fabric occurs in a separate section after cooling integrity is confirmed. The two domains are never combined in a single test step because one is internal copper and the other is external fiber.

What meter is required for the grounding resistance check?

A low-resistance ohmmeter capable of resolving milliohms is used between the rack ground bar and the facility electrode. The reading is recorded at both ends of the path and compared to the design target before any power is applied.

How are failed MPO strands handled in the ATP?

A failed strand triggers isolation of that trunk leg, re-cleaning or replacement, and a full retest of the affected link. The rack remains unpowered until the scale-out fabric passes its loss budget; internal NVLink copper tests are unaffected.

Who must sign each ATP section before the next rack begins?

Both the field lead and the customer commissioning engineer must initial every completed section. The document is archived with rack serial numbers only after all gates are satisfied.

What stops the ATP if a cooling pressure test fails?

Any measurable decay or visible moisture requires valve isolation, repair, and a complete retest of that loop segment. The rack cannot advance to electrical checks until the cooling system holds pressure for the full duration specified in the plan.

Ready to Deploy Your GPU Infrastructure?_

Tell us about your project. Book a call and we’ll discuss scope, timeline, and the best approach for your deployment.

Book a Call