Quiz: AI Data Center Reliability & Goodput (Part 12)

10 questions from Part 12 of the guide. Score instantly; every explanation links to the chapter that derives the answer.

The guide's core critique of legacy tier standards for AI clusters is that they…
ignore cooling entirelywere written for European gridsare too expensive to certifyrate a building, not a job — facility nines say nothing about GPU-hours converted to useful work
A service objective requires live maintenance and survival of the defined unplanned-fault case. What should the owner require before selecting a resilience topology?
select a Tier label and let the contractor infer the failure casesbuy the topology with the most installed equipmentconvert the desired outcome to a legacy availability percentageenumerate the maintenance and fault states, shared paths, transfer dependencies, and load behavior, then test the topology against them
Two resilience proposals quote different topology premiums. What normalization must happen before the owner can compare them?
compare installed nameplate alonemap the workload name directly to a topologyhold required IT load and service outcomes constant, then price installed capacity, operating policy, losses, and each maintenance and fault stateassume every spare asset is permanently stranded
Which stipulated training-goodput endpoints define the guide’s sensitivity case?
~90% vs ~96%~99% vs ~99.9%~75% vs ~85%~50% vs ~70%
On Meta's Llama 3 405B run on up to 16,384 H100s, how often did unexpected interruptions occur?
twice, totalabout 1 every 3 hours — 419 in 54 daysabout one per monthabout one per week
In the cited high-interruption training scenario, which recovery intervention has the largest modeled goodput leverage?
buying a Tier IV facilitymulti-tier checkpointing — Google reports checkpoint restore <1 min across clusters of thousands of nodesdoubling spare GPUsslowing training to reduce stress
At a training failure, what establishes the RPO boundary?
the storage vendor's replication SLAthe captured state of the newest valid checkpoint that survives the declared failure domain after durable commitzero — training must never lose workthe facility tier
Active-active geographic failover for an AI service costs roughly…
~20% extranothing if automated~2× — you carry a second live fleet~10% extra
A provider cites a 99% node / 95% rack SLA example. What can the buyer validly infer?
Only that the named example uses those thresholds; the buyer must compare measurement boundary, exclusions, credits, and goodput terms in each actual contractEvery provider with the same rack platform offers the same termsThose percentages are the 2026 market baselineThe rack will deliver at least 95% training goodput
What share of impactful data-center outages does power cause?
~45%~25%~10%~80%

Check your understanding: Reliability, Resilience & Standards