Quiz: AI Data Center Commissioning & Burn-In (Part 13)

The commissioning ladder the guide is built around is…

alpha, beta, GAL1–L5: factory tests → site arrival → pre-functional → functional → integrated systems testing of the whole facilitybronze to platinumthere is no standard ladder

Every commissioning script should carry…

a vendor signaturea hard, pre-agreed numerical acceptance gate with a named witness and redline-on-failureflexible targets to be fairphotographic evidence only

The guide's critique of static load-bank power acceptance is that…

load banks are too expensivesteady-state profiles never exercise the synchronized MW-scale transients an AI cluster actually producesload banks damage the UPSit takes too long

IST's black-building test proves…

the paint is drythe facility survives a full utility pull-the-plug — the cascade of ride-through, transfer, and generation actually choreographsthe network is fastbadge readers work

Fabric acceptance at the physical layer sets thresholds on…

cable colorspre-FEC/post-FEC bit-error rate, a link-flap budget, and a soak periodswitch firmware versions onlyport names

The cluster-level fabric gate before 'ready for production' is…

one clean iperf runa busbw floor per collective and message size — the fabric must deliver its effective bandwidth under real collective patternsvendor certificationcable certification alone

GPU burn-in soak windows run…

an afternoon72 hours minimum to 168 hours (7 days) for the full infant-mortality campaign30 days alwaysburn-in is obsolete

On a fresh cluster, what does the failure curve look like versus mature fleets?

identical from day onefar worse initially — expect 3–4 weeks of bring-up before rates flatten toward the ~7-day-per-512-GPU MTBF baselinebetter, since parts are newflat at zero for a year

The energization sequence decision is about…

which switch to flip first, ceremoniallyhow many blocks come up at once and whether each step preserves a safe retreat pathutility paperwork orderit's fully automated everywhere

The mechanical-Cx-to-burn-in boundary should be…

wherever schedules collidea clean handoff at full flow on load banks — cooling proven before GPUs become the test loadskipped to save timedecided by the GC

Check your understanding: Commissioning & Go-Live