NCA-AIIO - AI Infrastructure (40% of the exam) - Section 2.6
Identify facility requirements for accelerated infrastructure.
Identify the facility requirements that differentiate GPU-dense deployments from traditional server installations, including elevated power density per rack, specialised cooling infrastructure, floor load capacity, and redundant power distribution. Recognise how these requirements influence site selection and data centre design for AI workloads.
Practice question for this objective
lock_openFree sampleAI Infrastructuremedium
A facilities engineer is provisioning power for a row of GPU racks. Each rack draws a sustained load near its budgeted maximum, and the requirement is that the loss of any single power feed or distribution unit must not bring a rack down. Which power design principle most directly meets this resilience requirement?
ASizing a single power feed to exactly the rack's peak draw so that capacity is fully utilised and no headroom is wasted.
BRelying on the cluster scheduler to migrate jobs off a rack within seconds whenever a power feed reports a fault.
CLowering the per-GPU power cap until the whole rack fits within a fraction of one feed's rated capacity at all times.
DProviding redundant power feeds and distribution in an N+1 configuration so the rack keeps running if any one feed or distribution unit fails.check_circle Correct
Recognise that N+1 redundant power feeds and distribution protect a high-draw GPU rack against the failure of any single power component. N+1 redundancy means at least one more power feed or distribution unit is present than the rack actually needs, so when any single component fails the remaining capacity still covers the full sustained load and the rack stays powered.
Why A is wrong: It is tempting because tight sizing looks efficient, but a single feed with no headroom offers no resilience and would take the rack down the moment that feed fails.
Why B is wrong: It is tempting because schedulers can drain nodes, but software migration cannot prevent the sudden power loss of an unfed rack, so it does not satisfy a feed-redundancy requirement.
Why C is wrong: It is tempting because capping power reduces draw, but it sacrifices performance and still leaves the rack dependent on a single feed, so it does not provide redundancy.
Why D is correct: This is correct because N+1 redundancy supplies at least one spare feed or distribution path beyond the rack's need, so the loss of a single component does not interrupt the supply.
Answers that look right on this material and are not. Each one is a distractor from a different question in the NCA-AIIO bank for this domain.
N, where the installed power capacity exactly matches the load and there is no spare path to fall back on
errorWhy it is wrong: Tempting because N is the baseline capacity term, but N provides no redundancy at all, so losing a feed would drop the load and it cannot satisfy the requirement.
The raised-floor tile permeability rating to ensure adequate under-floor airflow for the rack row.
errorWhy it is wrong: Airflow permeability is a cooling concern, not a structural one, and does not address whether the floor slab can bear the concentrated point load of heavy GPU racks.
A single-phase domestic-style circuit feeding each rack from the nearest wall outlet in the room
errorWhy it is wrong: Tempting because it is the simplest wiring idea, but single-phase outlets cannot supply 40 kW per rack and offer no flexible tap-off along a row, so they are unsuitable for high-density GPU loads.