NCA - AI Infrastructure and Operations NCA-AIIO Practice Test 1
Question 1
The selected AI infrastructure safeguard is not implemented: Include interface, RDMA and switch telemetry for link health and congestion diagnosis. The observed situation is: A fabric design lists port speeds but omits any means of collecting error and retransmission counters. What exposure remains?
- An inappropriate default design can yield unpredictable low-latency workload behavior.
- Jobs can run slowly for days without a clue that link errors or retransmissions increased.
- Packet drops and uncontrolled pause behavior can cause severe tail latency in collective operations.
- Without coordinated feedback, bursts may cause rising latency and drops across the RDMA fabric.
Explanation
Correct Answer
B:The primary exposure is that jobs can run slowly for days without a clue that link errors or retransmissions increased. This follows from a fabric design lists port speeds but omits any means of collecting error and retransmission counters. where the omitted measure was to include interface, RDMA and switch telemetry for link health and congestion diagnosis. The technical reason is that reliable AI network operations need evidence beyond nominal link-up status or configured line rate. A good operational validation after repair would be: An injected link impairment produces a correlated alert and counters identifying the affected path.
Incorrect Answers
A:This exposure arises in a different situation: An engineer chooses between InfiniBand and Ethernet but assumes both have identical congestion and transport behavior. There, the corresponding safeguard is to document protocol-specific transport, management and performance requirements before selecting the AI fabric. In the present case, the more direct risk is that jobs can run slowly for days without a clue that link errors or retransmissions increased.
C:This exposure arises in a different situation: A team deploys RoCE traffic over standard congested Ethernet with no plan to handle packet loss or pauses. There, the corresponding safeguard is to design appropriate loss-control, priority-flow and congestion mechanisms for the approved RoCE fabric. In the present case, the more direct risk is that jobs can run slowly for days without a clue that link errors or retransmissions increased.
D:This exposure arises in a different situation: An AI cluster uses ECN-capable switches but network operators disable congestion feedback for RDMA flows. There, the corresponding safeguard is to enable supported explicit congestion notification and endpoint rate-response controls where required. In the present case, the more direct risk is that jobs can run slowly for days without a clue that link errors or retransmissions increased.
Question 2
A data-center engineer must resolve the following situation: A fabric design lists port speeds but omits any means of collecting error and retransmission counters. What is the MOST appropriate response?
- Inspect PFC pause counters and congestion telemetry before modifying quality-of-service configuration.
- Document protocol-specific transport, management and performance requirements before selecting the AI fabric.
- Include interface, RDMA and switch telemetry for link health and congestion diagnosis.
- Enable supported explicit congestion notification and endpoint rate-response controls where required.
Explanation
Correct Answer
C:The appropriate response is to include interface, RDMA and switch telemetry for link health and congestion diagnosis. This is specific to the case where a fabric design lists port speeds but omits any means of collecting error and retransmission counters. The mechanism is that reliable AI network operations need evidence beyond nominal link-up status or configured line rate. If ignored, jobs can run slowly for days without a clue that link errors or retransmissions increased. Validate the choice using this targeted evidence: An injected link impairment produces a correlated alert and counters identifying the affected path.
Incorrect Answers
A:This alternative addresses a different requirement: A storage and training fabric experiences repeated priority flow-control storms that slow unrelated traffic classes. It would be appropriate there to inspect PFC pause counters and congestion telemetry before modifying quality-of-service configuration. But the present scenario is about a fabric design lists port speeds but omits any means of collecting error and retransmission counters. For it, the needed control is to include interface, RDMA and switch telemetry for link health and congestion diagnosis.
B:This alternative addresses a different requirement: An engineer chooses between InfiniBand and Ethernet but assumes both have identical congestion and transport behavior. It would be appropriate there to document protocol-specific transport, management and performance requirements before selecting the AI fabric. But the present scenario is about a fabric design lists port speeds but omits any means of collecting error and retransmission counters. For it, the needed control is to include interface, RDMA and switch telemetry for link health and congestion diagnosis.
D:This alternative addresses a different requirement: An AI cluster uses ECN-capable switches but network operators disable congestion feedback for RDMA flows. It would be appropriate there to enable supported explicit congestion notification and endpoint rate-response controls where required. But the present scenario is about a fabric design lists port speeds but omits any means of collecting error and retransmission counters. For it, the needed control is to include interface, RDMA and switch telemetry for link health and congestion diagnosis.
Question 3
An architecture review proposes: Use a governed firmware baseline and controlled upgrade or rollback with node-level inventory comparison. The underlying requirement is: A service engineer upgrades GPU firmware on one node without recording it, leading to mismatched cluster behavior. What technical rationale MOST strongly supports it?
- Out-of-band management permits diagnosis of some host failures without relying on the broken operating system.
- Power-component health monitoring is needed even when workload throughput has not yet degraded.
- Consistent supported firmware and driver versions are part of cluster lifecycle management.
- Orderly maintenance requires coordination between facilities work and the active GPU job scheduler.
Explanation
Correct Answer
C:The governing principle is that consistent supported firmware and driver versions are part of cluster lifecycle management. It supports the measure to use a governed firmware baseline and controlled upgrade or rollback with node-level inventory comparison. because a service engineer upgrades GPU firmware on one node without recording it, leading to mismatched cluster behavior. Without the control, unreviewed version drift can create device instability or unexplained workload incompatibilities. The most useful acceptance evidence is: A compliance scan reports the outlier node and a controlled maintenance change restores the approved baseline.
Incorrect Answers
A:That principle is relevant to this distinct case: A GPU worker loses its primary operating-system network stack and technicians have no alternate management path. The associated response would be to provide secure out-of-band management access using supported BMC interfaces and separate network controls. It does not describe the central reason for this case, which is that consistent supported firmware and driver versions are part of cluster lifecycle management.
B:That principle is relevant to this distinct case: Data-center management sees correct fan speeds but receives no alert when an accelerator power module reports a fault. The associated response would be to integrate supported power-supply and chassis fault signals with prioritized alert routing. It does not describe the central reason for this case, which is that consistent supported firmware and driver versions are part of cluster lifecycle management.
D:That principle is relevant to this distinct case: Technicians take shared GPU nodes offline during business hours without coordinating active checkpointing jobs. The associated response would be to enforce maintenance windows with scheduler drain, workload notice and readiness verification. It does not describe the central reason for this case, which is that consistent supported firmware and driver versions are part of cluster lifecycle management.
Question 4
A GPU operations plan calls for: Apply visual defect detection with production-representative examples and inspection feedback. The reason is that a factory collects camera images of assembly lines to detect defective products before shipment. Which principle is MOST relevant?
- Manufacturing computer vision can support repeatable defect screening when lighting and product variation are represented.
- Simulation can expose AI systems to hazardous edge cases under controlled repeatable conditions.
- Forecasting predicts future numeric patterns and requires avoiding training on information not available at prediction time.
- Image segmentation predicts spatial regions, supporting detailed inspection rather than replacing clinical judgment.
Explanation
Correct Answer
A:The governing principle is that manufacturing computer vision can support repeatable defect screening when lighting and product variation are represented. It supports the measure to apply visual defect detection with production-representative examples and inspection feedback. because a factory collects camera images of assembly lines to detect defective products before shipment. Without the control, unexpected camera conditions may cause defective goods to pass unchecked. The most useful acceptance evidence is: A production-line trial records missed defects and false rejects across lighting and product variants.
Incorrect Answers
B:That principle is relevant to this distinct case: An autonomous vehicle company trains a vision policy and needs rare-road-condition examples without public-road risk. The associated response would be to use high-fidelity simulation or synthetic data generation with realism and generalization testing. It does not describe the central reason for this case, which is that manufacturing computer vision can support repeatable defect screening when lighting and product variation are represented.
C:That principle is relevant to this distinct case: An energy provider forecasts electricity demand from time-series usage, calendar and weather data. The associated response would be to select a time-series forecasting approach and test prediction error over future held-out periods. It does not describe the central reason for this case, which is that manufacturing computer vision can support repeatable defect screening when lighting and product variation are represented.
D:That principle is relevant to this distinct case: A radiology team needs to highlight suspicious structures in diagnostic images for qualified human review. The associated response would be to evaluate computer-vision segmentation with clinically governed validation and human oversight. It does not describe the central reason for this case, which is that manufacturing computer vision can support repeatable defect screening when lighting and product variation are represented.
Question 5
In this situation, a model team trains on datasets whose source and consent status are not recorded across experiments. The team does not establish versioned data ingestion, lineage and access controls before model-development experiments. Which consequence BEST characterizes the gap?
- Retraining may become irreproducible or rely on unauthorized data that cannot be audited.
- A faulty input batch can cause an inferior model to replace a stable production version.
- Repeated test-set tuning can exaggerate real-world quality and lead to unsafe deployment decisions.
- The team may be unable to restore reliable predictions quickly after a problematic release.
Explanation
Correct Answer
A:The primary exposure is that retraining may become irreproducible or rely on unauthorized data that cannot be audited. This follows from a model team trains on datasets whose source and consent status are not recorded across experiments. where the omitted measure was to establish versioned data ingestion, lineage and access controls before model-development experiments. The technical reason is that aI lifecycles begin with governed data preparation and traceable inputs, not just accelerator allocation. A good operational validation after repair would be: A sampled training dataset traces back to approved origin, version, transformation and access authorization.
Incorrect Answers
B:This exposure arises in a different situation: An organization automatically retrains a model nightly and publishes weights without independent acceptance rules. There, the corresponding safeguard is to gate retraining output on dataset validation, safety metrics and approved promotion criteria. In the present case, the more direct risk is that retraining may become irreproducible or rely on unauthorized data that cannot be audited.
C:This exposure arises in a different situation: A team reports model accuracy using the same data that were repeatedly used for hyperparameter selection. There, the corresponding safeguard is to keep an appropriately isolated final test set and document evaluation metrics before selecting a release. In the present case, the more direct risk is that retraining may become irreproducible or rely on unauthorized data that cannot be audited.
D:This exposure arises in a different situation: A newly promoted inference model produces biased results and no earlier artifact or configuration was retained. There, the corresponding safeguard is to maintain a tested rollback path with previous model artifacts, endpoints and compatibility metadata. In the present case, the more direct risk is that retraining may become irreproducible or rely on unauthorized data that cannot be audited.
Question 6
The team recommends this safeguard: Evaluate data-parallel workers with synchronized gradients and a suitable global batch strategy. It applies because a model fits on one GPU but an organization wants to speed up training using more GPUs on the same dataset. Which concept justifies the design?
- Data parallelism replicates the model and partitions data while coordinating parameter updates.
- Adding more distributed nodes expands the operational failure surface as well as compute capacity.
- GPU partitioning is a workload-isolation choice, not an automatic substitute for full-device training capacity.
- Scale-up and scale-out use different interconnect domains and communication constraints.
Explanation
Correct Answer
A:The governing principle is that data parallelism replicates the model and partitions data while coordinating parameter updates. It supports the measure to evaluate data-parallel workers with synchronized gradients and a suitable global batch strategy. because a model fits on one GPU but an organization wants to speed up training using more GPUs on the same dataset. Without the control, naive replication without synchronization can produce divergent models or misleading throughput. The most useful acceptance evidence is: A multi-GPU run converges consistently and records useful samples per second and sync overhead.
Incorrect Answers
B:That principle is relevant to this distinct case: A GPU cluster loses progress whenever a worker disappears during a large multi-node training epoch. The associated response would be to plan checkpoint intervals, scheduler restart rules and failure-domain containment for distributed jobs. It does not describe the central reason for this case, which is that data parallelism replicates the model and partitions data while coordinating parameter updates.
C:That principle is relevant to this distinct case: A multi-GPU training model requires full-GPU memory but a manager proposes partitioning every GPU into tiny MIG slices. The associated response would be to reserve suitable full GPU devices for workloads that need whole-device capacity or communication topology. It does not describe the central reason for this case, which is that data parallelism replicates the model and partitions data while coordinating parameter updates.
D:That principle is relevant to this distinct case: A team adds more GPUs inside one system and assumes the same link topology extends to GPUs in other racks. The associated response would be to distinguish node-local NVLink or NVSwitch scaling from networked multi-node scale-out design. It does not describe the central reason for this case, which is that data parallelism replicates the model and partitions data while coordinating parameter updates.
Question 7
The following safeguard was deployed: Enable supported explicit congestion notification and endpoint rate-response controls where required. What operational observation would demonstrate its success?
- A controlled load test confirms correct ECN or PFC behavior and stable RDMA throughput.
- A traffic burst demonstrates congestion signaling and controlled send rates with acceptable latency.
- A design review names the actual transport and validates throughput and congestion settings in a pilot.
- An approved end-to-end path test passes the required packet sizes across all intended hops.
Explanation
Correct Answer
B:The strongest proof is: A traffic burst demonstrates congestion signaling and controlled send rates with acceptable latency. It directly tests the safeguard to enable supported explicit congestion notification and endpoint rate-response controls where required. The control addresses this specific context: An AI cluster uses ECN-capable switches but network operators disable congestion feedback for RDMA flows. It matters because eCN conveys emerging network congestion so senders can reduce demand before severe queue buildup. Without it, without coordinated feedback, bursts may cause rising latency and drops across the RDMA fabric.
Incorrect Answers
A:This observation would validate a different control: Design appropriate loss-control, priority-flow and congestion mechanisms for the approved RoCE fabric. It fits the case where a team deploys RoCE traffic over standard congested Ethernet with no plan to handle packet loss or pauses. It cannot prove success here; instead the required evidence is a traffic burst demonstrates congestion signaling and controlled send rates with acceptable latency.
C:This observation would validate a different control: Document protocol-specific transport, management and performance requirements before selecting the AI fabric. It fits the case where an engineer chooses between InfiniBand and Ethernet but assumes both have identical congestion and transport behavior. It cannot prove success here; instead the required evidence is a traffic burst demonstrates congestion signaling and controlled send rates with acceptable latency.
D:This observation would validate a different control: Validate path MTU consistency across endpoints, switches and relevant routed interfaces. It fits the case where a team sets jumbo-frame MTU on GPU servers but overlooks one switch hop in a high-speed data path. It cannot prove success here; instead the required evidence is a traffic burst demonstrates congestion signaling and controlled send rates with acceptable latency.
Question 8
The proposed approach is to monitor used and available GPU memory across realistic serving contexts and concurrent workload sizes. In the following setting, a model endpoint experiences out-of-memory failures only during long prompts and concurrent sessions. Which principle BEST supports the approach?
- High activity percentage does not rule out throughput loss from thermal limits.
- GPU memory telemetry must capture runtime allocations and usage peaks rather than only static model size.
- Corrected memory errors can provide early hardware-risk evidence even before visible job failure.
- Xid diagnostics provide context for GPU software or hardware failures and require supported interpretation.
Explanation
Correct Answer
B:The governing principle is that gPU memory telemetry must capture runtime allocations and usage peaks rather than only static model size. It supports the measure to monitor used and available GPU memory across realistic serving contexts and concurrent workload sizes. because a model endpoint experiences out-of-memory failures only during long prompts and concurrent sessions. Without the control, the service can fail during demand peaks although idle memory looks sufficient. The most useful acceptance evidence is: A high-context load test reports peak memory and triggers a controlled alert before allocation failure.
Incorrect Answers
A:That principle is relevant to this distinct case: An accelerator runs at reduced clock speed during a full training epoch but device utilization appears high. The associated response would be to correlate temperature, clock state and thermal-throttling reasons during sustained workload measurements. It does not describe the central reason for this case, which is that gPU memory telemetry must capture runtime allocations and usage peaks rather than only static model size.
C:That principle is relevant to this distinct case: A GPU node has increasing corrected memory errors but operations waits for application crashes before investigating. The associated response would be to collect supported ECC error counters and evaluate hardware-health escalation thresholds. It does not describe the central reason for this case, which is that gPU memory telemetry must capture runtime allocations and usage peaks rather than only static model size.
D:That principle is relevant to this distinct case: A containerized training job unexpectedly exits and system logs contain GPU Xid error messages. The associated response would be to correlate NVIDIA Xid events with driver status, hardware health and the affected job timeline. It does not describe the central reason for this case, which is that gPU memory telemetry must capture runtime allocations and usage peaks rather than only static model size.
Question 9
An organization is evaluating an AI deployment: A data center scales from four to sixty-four training nodes without upgrading per-node network ports. What solution MOST directly fits the stated need?
- Define capacity headroom and controlled drain/replacement procedures based on job availability targets.
- Evaluate data-parallel workers with synchronized gradients and a suitable global batch strategy.
- Profile collective operations and locate the straggling node, link or synchronization pattern.
- Model aggregate east-west traffic and network oversubscription at the intended cluster scale.
Explanation
Correct Answer
D:The appropriate response is to model aggregate east-west traffic and network oversubscription at the intended cluster scale. This is specific to the case where a data center scales from four to sixty-four training nodes without upgrading per-node network ports. The mechanism is that distributed accelerator capacity cannot scale efficiently when network bisection bandwidth is insufficient. If ignored, the new fleet may deliver little additional model throughput while collective traffic saturates links. Validate the choice using this targeted evidence: A scale test plots iteration throughput and switch utilization at increasing node counts.
Incorrect Answers
A:This alternative addresses a different requirement: A research group buys all available GPUs with no spare capacity for maintenance or failed accelerator replacement. It would be appropriate there to define capacity headroom and controlled drain/replacement procedures based on job availability targets. But the present scenario is about a data center scales from four to sixty-four training nodes without upgrading per-node network ports. For it, the needed control is to model aggregate east-west traffic and network oversubscription at the intended cluster scale.
B:This alternative addresses a different requirement: A model fits on one GPU but an organization wants to speed up training using more GPUs on the same dataset. It would be appropriate there to evaluate data-parallel workers with synchronized gradients and a suitable global batch strategy. But the present scenario is about a data center scales from four to sixty-four training nodes without upgrading per-node network ports. For it, the needed control is to model aggregate east-west traffic and network oversubscription at the intended cluster scale.
C:This alternative addresses a different requirement: A distributed training run uses 32 GPUs but each training step waits for one slow all-reduce participant. It would be appropriate there to profile collective operations and locate the straggling node, link or synchronization pattern. But the present scenario is about a data center scales from four to sixty-four training nodes without upgrading per-node network ports. For it, the needed control is to model aggregate east-west traffic and network oversubscription at the intended cluster scale.
Question 10
A team omits this measure: Plan checkpoint intervals, scheduler restart rules and failure-domain containment for distributed jobs. In this environment: A GPU cluster loses progress whenever a worker disappears during a large multi-node training epoch. Which resulting problem is MOST likely?
- A transient node outage can waste all work since the last recoverable coordinated checkpoint.
- Expensive devices can sit idle while requests are stranded behind bad resource placement.
- A single overloaded network link or faulty participant can eliminate expected scaling gains.
- Naive replication without synchronization can produce divergent models or misleading throughput.
Explanation
Correct Answer
A:The primary exposure is that a transient node outage can waste all work since the last recoverable coordinated checkpoint. This follows from a GPU cluster loses progress whenever a worker disappears during a large multi-node training epoch. where the omitted measure was to plan checkpoint intervals, scheduler restart rules and failure-domain containment for distributed jobs. The technical reason is that adding more distributed nodes expands the operational failure surface as well as compute capacity. A good operational validation after repair would be: An authorized node-failure exercise resumes from a valid checkpoint without corrupt model state.
Incorrect Answers
B:This exposure arises in a different situation: A shared accelerator fleet has idle GPUs in some nodes while large training jobs wait without proper scheduling. There, the corresponding safeguard is to use an appropriate workload scheduler with GPU resource accounting and topology-aware placement. In the present case, the more direct risk is that a transient node outage can waste all work since the last recoverable coordinated checkpoint.
C:This exposure arises in a different situation: A distributed training run uses 32 GPUs but each training step waits for one slow all-reduce participant. There, the corresponding safeguard is to profile collective operations and locate the straggling node, link or synchronization pattern. In the present case, the more direct risk is that a transient node outage can waste all work since the last recoverable coordinated checkpoint.
D:This exposure arises in a different situation: A model fits on one GPU but an organization wants to speed up training using more GPUs on the same dataset. There, the corresponding safeguard is to evaluate data-parallel workers with synchronized gradients and a suitable global batch strategy. In the present case, the more direct risk is that a transient node outage can waste all work since the last recoverable coordinated checkpoint.