NCA-AIIO - AI Infrastructure - Section 2.5

Identify key components and considerations of accelerated infrastructure clusters.

Identify the key components of an accelerated computing cluster, including GPU servers, high-speed compute fabric, storage systems, and out-of-band management networks. Recognise considerations such as GPU-to-GPU interconnect bandwidth, storage I/O throughput, and network topology that affect overall cluster performance for AI workloads.

Practice question for this objective

Free sampleAI Infrastructuremedium

An architect is reviewing the network design for a new accelerated cluster. The plan places GPU collective traffic, parallel-storage reads and out-of-band management telemetry all on one shared fabric to simplify cabling. During trial runs the team sees gradient all-reduce throughput collapse whenever a large checkpoint is written. Which design principle should the architect apply to prevent this interference?

  • ARaise the management network bandwidth so that telemetry no longer competes with the storage and compute traffic on the shared fabric.
  • BSchedule all checkpoint writes to occur only at the very end of each training job so they never overlap with gradient exchange.
  • CSwitch the shared fabric from InfiniBand to Ethernet so that congestion control on Ethernet automatically resolves the contention between flows.
  • DSeparate the cluster onto purpose-built fabrics, giving GPU collective traffic its own dedicated compute fabric isolated from storage and management networks. Correct
Recognise that accelerated clusters segregate compute, storage and management traffic onto dedicated fabrics to stop storage bursts from throttling GPU collectives. Reference accelerated-cluster architectures place GPU collective traffic on a dedicated compute fabric, separate from the storage and management networks, so that high-volume storage events such as checkpoint writes cannot consume the bandwidth that latency-sensitive all-reduce operations depend on.

Why A is wrong: It is tempting because more bandwidth sounds like it would relieve congestion, but the contention is between storage and compute flows, so enlarging the low-volume management network does not stop checkpoints from throttling all-reduce.

Why B is wrong: It is tempting because timing the writes avoids overlap, but periodic checkpointing during long runs is needed for fault tolerance, so deferring it sacrifices resilience and still leaves the fabrics coupled.

Why C is wrong: It is tempting because Ethernet has congestion-control features, but changing the link technology on a single shared fabric does not separate the competing flows, and the storage bursts would still degrade collectives.

Why D is correct: This is correct because reference accelerated-cluster designs segregate traffic by purpose so that bursty storage writes cannot steal bandwidth from latency-sensitive GPU collective operations.

See more NCA-AIIO practice questions, answers explained.

More in this domain

Back to all AI Infrastructure objectives, or the NCA-AIIO cert hub.

Examworthy is not affiliated with or endorsed by NVIDIA. Original, blueprint-aligned practice material only.