NCA - AI Infrastructure and Operations NCA-AIIO Practice Test 21
Question 200
The design team faces this case: A contact center wants automatic transcription of customer calls for searchable support records. It adopts: Use automatic speech recognition with language coverage, privacy controls and transcription quality checks. What evidence BEST validates the intended outcome?
- Backtesting uses chronologically separated windows and reports error across season and demand levels.
- Evaluation on consented noisy calls reports word error rate and appropriate retention controls.
- Held-out images are scored against expert annotations and a reviewer workflow remains mandatory.
- An offline ranking benchmark and controlled experiment compare relevance and business outcomes.
Explanation
Correct Answer
B:The strongest proof is: Evaluation on consented noisy calls reports word error rate and appropriate retention controls. It directly tests the safeguard to use automatic speech recognition with language coverage, privacy controls and transcription quality checks. The control addresses this specific context: A contact center wants automatic transcription of customer calls for searchable support records. It matters because speech recognition converts audio to text and must account for noise, accents and domain-specific terms. Without it, transcript errors can misrepresent customer instructions while sensitive recordings remain exposed.
Incorrect Answers
A:This observation would validate a different control: Select a time-series forecasting approach and test prediction error over future held-out periods. It fits the case where an energy provider forecasts electricity demand from time-series usage, calendar and weather data. It cannot prove success here; instead the required evidence is evaluation on consented noisy calls reports word error rate and appropriate retention controls.
C:This observation would validate a different control: Evaluate computer-vision segmentation with clinically governed validation and human oversight. It fits the case where a radiology team needs to highlight suspicious structures in diagnostic images for qualified human review. It cannot prove success here; instead the required evidence is evaluation on consented noisy calls reports word error rate and appropriate retention controls.
D:This observation would validate a different control: Consider a recommendation model trained on consented interaction data and evaluate ranking quality. It fits the case where a retailer wants to recommend products based on customers' prior interactions and item characteristics. It cannot prove success here; instead the required evidence is evaluation on consented noisy calls reports word error rate and appropriate retention controls.
Question 201
A data-center manager declines to install and validate supported GPU drivers, NVIDIA device integration and resource advertisement. The operational context is: A Kubernetes AI cluster contains NVIDIA GPUs but pod resource requests are invisible to the control plane. Which risk is MOST direct?
- A model needing a specific GPU feature could be sent to an incompatible node.
- GPU pods may remain pending or start without usable devices even when hardware exists.
- Priority enforcement can waste GPU hours when jobs cannot resume from checkpoint state.
- One tenant can monopolize GPUs and starve urgent inference or research commitments.
Explanation
Correct Answer
B:The primary exposure is that gPU pods may remain pending or start without usable devices even when hardware exists. This follows from a Kubernetes AI cluster contains NVIDIA GPUs but pod resource requests are invisible to the control plane. where the omitted measure was to install and validate supported GPU drivers, NVIDIA device integration and resource advertisement. The technical reason is that kubernetes requires an appropriate device discovery and allocation path to schedule accelerator workloads. A good operational validation after repair would be: An authorized test pod requests one advertised GPU and can execute the approved diagnostic workload.
Incorrect Answers
A:This exposure arises in a different situation: Kubernetes workers contain different GPU types but all nodes are treated as equivalent for AI job placement. There, the corresponding safeguard is to use hardware feature discovery and node labels or resource constraints for workload-specific placement. In the present case, the more direct risk is that gPU pods may remain pending or start without usable devices even when hardware exists.
C:This exposure arises in a different situation: Training jobs are repeatedly preempted by higher-priority work but restart from scratch because the scheduler lacks recovery hooks. There, the corresponding safeguard is to coordinate preemption policy with checkpoint and requeue support for eligible training jobs. In the present case, the more direct risk is that gPU pods may remain pending or start without usable devices even when hardware exists.
D:This exposure arises in a different situation: Multiple research teams compete for a shared GPU fleet and one account submits hundreds of low-priority jobs. There, the corresponding safeguard is to define resource quotas, scheduling priorities and fair-share policy with accountable ownership. In the present case, the more direct risk is that gPU pods may remain pending or start without usable devices even when hardware exists.
Question 202
The team recommends this safeguard: Select an appropriate high-bandwidth GPU interconnect topology for the planned multi-GPU communication. It applies because a deep-learning training server contains several GPUs but inter-GPU copies must cross a slower host path. Which concept justifies the design?
- Reliable training infrastructure must store and restore evolving state as well as original training data.
- Inter-GPU bandwidth and topology affect scaling when tensors are exchanged frequently.
- AI training input pipelines must deliver data at the rate accelerators can consume it.
- Training working sets include intermediate and optimizer state beyond the model's serialized weight file.
Explanation
Correct Answer
B:The governing principle is that inter-GPU bandwidth and topology affect scaling when tensors are exchanged frequently. It supports the measure to select an appropriate high-bandwidth GPU interconnect topology for the planned multi-GPU communication. because a deep-learning training server contains several GPUs but inter-GPU copies must cross a slower host path. Without the control, distributed training may spend most of its time moving data instead of computing gradients. The most useful acceptance evidence is: A topology-aware bandwidth test and training profile show acceptable GPU-to-GPU transfer time.
Incorrect Answers
A:That principle is relevant to this distinct case: A training program uses frequent multi-terabyte checkpoints but reserves storage for only its original dataset. The associated response would be to include checkpoint size, frequency, retention and restore throughput in capacity and storage planning. It does not describe the central reason for this case, which is that inter-GPU bandwidth and topology affect scaling when tensors are exchanged frequently.
C:That principle is relevant to this distinct case: A vision training job reads millions of small images from one slow network file share and leaves accelerators waiting. The associated response would be to plan storage throughput, metadata operations and prefetch behavior using representative input data. It does not describe the central reason for this case, which is that inter-GPU bandwidth and topology affect scaling when tensors are exchanged frequently.
D:That principle is relevant to this distinct case: A training model needs 52 GB of weights and activations but the proposed accelerator has only 24 GB of usable memory. The associated response would be to size accelerator memory for model parameters, activations, gradients and runtime headroom before procurement. It does not describe the central reason for this case, which is that inter-GPU bandwidth and topology affect scaling when tensors are exchanged frequently.
Question 203
An operations team implements: Compare elastic cloud accelerator rental against on-premises ownership using realistic utilization assumptions. Which test result would MOST convincingly show the approach works?
- Capacity and interruption tests show serving stays available while resumable training tolerates a reclaimed worker.
- A total-cost model accounts for usage hours, staffing, storage, network transfer and capital depreciation.
- The proposed operating model assigns trained owners and budgets recurring support and facility work.
- A controlled promotion records model identity and proves inference receives only permitted data in each region.
Explanation
Correct Answer
B:The strongest proof is: A total-cost model accounts for usage hours, staffing, storage, network transfer and capital depreciation. It directly tests the safeguard to compare elastic cloud accelerator rental against on-premises ownership using realistic utilization assumptions. The control addresses this specific context: A small research lab runs GPU training only six days per quarter and is deciding whether to purchase a large fleet. It matters because low-duty-cycle workloads often favor flexible capacity, while high continuous utilization can favor owned infrastructure. Without it, an idle owned cluster can impose large fixed costs despite infrequent useful training.
Incorrect Answers
A:This observation would validate a different control: Select stable serving capacity and an appropriate lower-cost interruptible training tier with checkpoint recovery. It fits the case where an AI service has predictable daily demand but a training task can be interrupted and resumed later. It cannot prove success here; instead the required evidence is a total-cost model accounts for usage hours, staffing, storage, network transfer and capital depreciation.
C:This observation would validate a different control: Include operational staffing and maintenance lifecycle in the on-premises versus managed-cloud decision. It fits the case where a startup wants a fully owned GPU platform but lacks specialists to operate switching, cooling and firmware updates. It cannot prove success here; instead the required evidence is a total-cost model accounts for usage hours, staffing, storage, network transfer and capital depreciation.
D:This observation would validate a different control: Design a hybrid model lifecycle with approved artifact replication and separate sensitive-data access rules. It fits the case where a regulated enterprise trains models locally but wants cloud inference for geographically distributed customers. It cannot prove success here; instead the required evidence is a total-cost model accounts for usage hours, staffing, storage, network transfer and capital depreciation.
Question 204
An AI infrastructure team faces this requirement: An operations team tracks training-job completion but has no aggregate view of failing nodes or rack utilization. Which action BEST addresses the operational constraint?
- Create an owned incident workflow with timeline, affected device identity, evidence and escalation rules.
- Enforce maintenance windows with scheduler drain, workload notice and readiness verification.
- Track usable versus installed GPU hours, allocation, job backlog and maintenance losses before purchasing.
- Build a central fleet dashboard joining node inventory, job states, health and GPU-capacity measures.
Explanation
Correct Answer
D:The appropriate response is to build a central fleet dashboard joining node inventory, job states, health and GPU-capacity measures. This is specific to the case where an operations team tracks training-job completion but has no aggregate view of failing nodes or rack utilization. The mechanism is that fleet operations need correlated resource and job visibility to distinguish isolated failures from broader constraints. If ignored, repeated accelerator faults or stranded capacity can stay hidden among individual job tickets. Validate the choice using this targeted evidence: The fleet dashboard identifies the affected rack and correlates its degraded nodes with failed job records.
Incorrect Answers
A:This alternative addresses a different requirement: A node reports intermittent GPU failures but operators simply restart the job without recording the incident context. It would be appropriate there to create an owned incident workflow with timeline, affected device identity, evidence and escalation rules. But the present scenario is about an operations team tracks training-job completion but has no aggregate view of failing nodes or rack utilization. For it, the needed control is to build a central fleet dashboard joining node inventory, job states, health and GPU-capacity measures.
B:This alternative addresses a different requirement: Technicians take shared GPU nodes offline during business hours without coordinating active checkpointing jobs. It would be appropriate there to enforce maintenance windows with scheduler drain, workload notice and readiness verification. But the present scenario is about an operations team tracks training-job completion but has no aggregate view of failing nodes or rack utilization. For it, the needed control is to build a central fleet dashboard joining node inventory, job states, health and GPU-capacity measures.
C:This alternative addresses a different requirement: A research group requests another rack of GPUs while existing capacity is partly unavailable or poorly scheduled. It would be appropriate there to track usable versus installed GPU hours, allocation, job backlog and maintenance losses before purchasing. But the present scenario is about an operations team tracks training-job completion but has no aggregate view of failing nodes or rack utilization. For it, the needed control is to build a central fleet dashboard joining node inventory, job states, health and GPU-capacity measures.
Question 205
The following safeguard was deployed: Profile collective operations and locate the straggling node, link or synchronization pattern. What operational observation would demonstrate its success?
- A maintenance rehearsal removes a GPU node while priority jobs meet acceptable queue and recovery targets.
- Per-rank timing identifies the slow participant and a controlled repair reduces collective duration.
- A scale test plots iteration throughput and switch utilization at increasing node counts.
- A topology diagram and scale-out test show the actual network path and collective performance.
Explanation
Correct Answer
B:The strongest proof is: Per-rank timing identifies the slow participant and a controlled repair reduces collective duration. It directly tests the safeguard to profile collective operations and locate the straggling node, link or synchronization pattern. The control addresses this specific context: A distributed training run uses 32 GPUs but each training step waits for one slow all-reduce participant. It matters because collective communication time can define total iteration duration across synchronized workers. Without it, a single overloaded network link or faulty participant can eliminate expected scaling gains.
Incorrect Answers
A:This observation would validate a different control: Define capacity headroom and controlled drain/replacement procedures based on job availability targets. It fits the case where a research group buys all available GPUs with no spare capacity for maintenance or failed accelerator replacement. It cannot prove success here; instead the required evidence is per-rank timing identifies the slow participant and a controlled repair reduces collective duration.
C:This observation would validate a different control: Model aggregate east-west traffic and network oversubscription at the intended cluster scale. It fits the case where a data center scales from four to sixty-four training nodes without upgrading per-node network ports. It cannot prove success here; instead the required evidence is per-rank timing identifies the slow participant and a controlled repair reduces collective duration.
D:This observation would validate a different control: Distinguish node-local NVLink or NVSwitch scaling from networked multi-node scale-out design. It fits the case where a team adds more GPUs inside one system and assumes the same link topology extends to GPUs in other racks. It cannot prove success here; instead the required evidence is per-rank timing identifies the slow participant and a controlled repair reduces collective duration.
Question 206
An AI infrastructure team faces this requirement: A team submits two tightly coupled distributed training pods but one starts long before resources for the other are ready. Which action BEST addresses the operational constraint?
- Cordon or drain the node through approved scheduler workflows and coordinate checkpointable tasks.
- Use gang-style scheduling or coordinated job admission where supported for dependent distributed workers.
- Enforce supported device allocation, tenant permissions and container isolation for GPU workloads.
- Define resource quotas, scheduling priorities and fair-share policy with accountable ownership.
Explanation
Correct Answer
B:The appropriate response is to use gang-style scheduling or coordinated job admission where supported for dependent distributed workers. This is specific to the case where a team submits two tightly coupled distributed training pods but one starts long before resources for the other are ready. The mechanism is that synchronized training workers often need coordinated placement to avoid holding idle resources. If ignored, partially admitted distributed jobs can occupy GPUs while making no training progress. Validate the choice using this targeted evidence: A resource shortage test delays the job group until the required worker count can start together.
Incorrect Answers
A:This alternative addresses a different requirement: A GPU node is marked for maintenance while the scheduler continues assigning new long-running training jobs to it. It would be appropriate there to cordon or drain the node through approved scheduler workflows and coordinate checkpointable tasks. But the present scenario is about a team submits two tightly coupled distributed training pods but one starts long before resources for the other are ready. For it, the needed control is to use gang-style scheduling or coordinated job admission where supported for dependent distributed workers.
C:This alternative addresses a different requirement: An inference container requests a GPU but can access every accelerator and shared host filesystem on its node. It would be appropriate there to enforce supported device allocation, tenant permissions and container isolation for GPU workloads. But the present scenario is about a team submits two tightly coupled distributed training pods but one starts long before resources for the other are ready. For it, the needed control is to use gang-style scheduling or coordinated job admission where supported for dependent distributed workers.
D:This alternative addresses a different requirement: Multiple research teams compete for a shared GPU fleet and one account submits hundreds of low-priority jobs. It would be appropriate there to define resource quotas, scheduling priorities and fair-share policy with accountable ownership. But the present scenario is about a team submits two tightly coupled distributed training pods but one starts long before resources for the other are ready. For it, the needed control is to use gang-style scheduling or coordinated job admission where supported for dependent distributed workers.
Question 207
An architecture review proposes: Assess supported DPU offload for suitable networking and storage functions and measure actual net benefit. The underlying requirement is: A busy accelerated server spends significant host CPU cycles processing storage and network control tasks. What technical rationale MOST strongly supports it?
- DPU deployment adds operational elements that need observability alongside GPU and host status.
- Dedicated infrastructure processing can separate selected network and security duties from general host code.
- DPUs specialize in infrastructure data-plane operations and do not replace GPUs for AI model arithmetic.
- DPUs can execute selected infrastructure functions on dedicated resources, reducing host CPU contention.
Explanation
Correct Answer
D:The governing principle is that dPUs can execute selected infrastructure functions on dedicated resources, reducing host CPU contention. It supports the measure to assess supported DPU offload for suitable networking and storage functions and measure actual net benefit. because a busy accelerated server spends significant host CPU cycles processing storage and network control tasks. Without the control, the host may lose useful application cycles when all infrastructure handling shares CPU cores. The most useful acceptance evidence is: A controlled comparison shows lower host CPU demand without reducing storage or network correctness.
Incorrect Answers
A:That principle is relevant to this distinct case: A cluster sees intermittent network errors but its DPU ports and offload services are missing from health dashboards. The associated response would be to monitor supported BlueField interfaces, service state and relevant packet or error counters. It does not describe the central reason for this case, which is that dPUs can execute selected infrastructure functions on dedicated resources, reducing host CPU contention.
B:That principle is relevant to this distinct case: A security architect wants infrastructure service isolation from host applications that may be compromised. The associated response would be to evaluate BlueField DPU-supported trusted infrastructure processing and its administrative trust boundaries. It does not describe the central reason for this case, which is that dPUs can execute selected infrastructure functions on dedicated resources, reducing host CPU contention.
C:That principle is relevant to this distinct case: A platform team installs a DPU and assumes it will automatically accelerate every CUDA compute kernel. The associated response would be to assign appropriate network, security or storage infrastructure tasks to the DPU rather than model tensor computation. It does not describe the central reason for this case, which is that dPUs can execute selected infrastructure functions on dedicated resources, reducing host CPU contention.
Question 208
An AI infrastructure team faces this requirement: A vision training job reads millions of small images from one slow network file share and leaves accelerators waiting. Which action BEST addresses the operational constraint?
- Size accelerator memory for model parameters, activations, gradients and runtime headroom before procurement.
- Include checkpoint size, frequency, retention and restore throughput in capacity and storage planning.
- Confirm accelerator data-type support and validate precision choice against convergence requirements.
- Plan storage throughput, metadata operations and prefetch behavior using representative input data.
Explanation
Correct Answer
D:The appropriate response is to plan storage throughput, metadata operations and prefetch behavior using representative input data. This is specific to the case where a vision training job reads millions of small images from one slow network file share and leaves accelerators waiting. The mechanism is that aI training input pipelines must deliver data at the rate accelerators can consume it. If ignored, gPU resources remain expensive and idle when a storage system cannot supply enough samples. Validate the choice using this targeted evidence: A full input-path benchmark records steady sample feed rate and GPU wait time under concurrent training.
Incorrect Answers
A:This alternative addresses a different requirement: A training model needs 52 GB of weights and activations but the proposed accelerator has only 24 GB of usable memory. It would be appropriate there to size accelerator memory for model parameters, activations, gradients and runtime headroom before procurement. But the present scenario is about a vision training job reads millions of small images from one slow network file share and leaves accelerators waiting. For it, the needed control is to plan storage throughput, metadata operations and prefetch behavior using representative input data.
B:This alternative addresses a different requirement: A training program uses frequent multi-terabyte checkpoints but reserves storage for only its original dataset. It would be appropriate there to include checkpoint size, frequency, retention and restore throughput in capacity and storage planning. But the present scenario is about a vision training job reads millions of small images from one slow network file share and leaves accelerators waiting. For it, the needed control is to plan storage throughput, metadata operations and prefetch behavior using representative input data.
C:This alternative addresses a different requirement: A training team buys accelerators rated for maximum low-precision throughput while requiring strict numeric convergence. It would be appropriate there to confirm accelerator data-type support and validate precision choice against convergence requirements. But the present scenario is about a vision training job reads millions of small images from one slow network file share and leaves accelerators waiting. For it, the needed control is to plan storage throughput, metadata operations and prefetch behavior using representative input data.
Question 209
An architecture review proposes: Define resource quotas, scheduling priorities and fair-share policy with accountable ownership. The underlying requirement is: Multiple research teams compete for a shared GPU fleet and one account submits hundreds of low-priority jobs. What technical rationale MOST strongly supports it?
- Multi-tenant accelerator scheduling must balance fairness and business priority rather than first-come volume alone.
- Workload admission policies should protect latency-sensitive serving from batch training contention.
- A cluster scheduler should match interruption rules to each workload's recoverability characteristics.
- Job schedulers must understand accelerator resources rather than treating all CPU nodes as interchangeable.
Explanation
Correct Answer
A:The governing principle is that multi-tenant accelerator scheduling must balance fairness and business priority rather than first-come volume alone. It supports the measure to define resource quotas, scheduling priorities and fair-share policy with accountable ownership. because multiple research teams compete for a shared GPU fleet and one account submits hundreds of low-priority jobs. Without the control, one tenant can monopolize GPUs and starve urgent inference or research commitments. The most useful acceptance evidence is: An admission test respects tenant quotas while an authorized high-priority job receives agreed capacity.
Incorrect Answers
B:That principle is relevant to this distinct case: An AI platform runs model-serving jobs and distributed training under one global namespace with no capacity boundaries. The associated response would be to separate workload classes with appropriate quotas, priorities and isolation rules in the scheduler. It does not describe the central reason for this case, which is that multi-tenant accelerator scheduling must balance fairness and business priority rather than first-come volume alone.
C:That principle is relevant to this distinct case: Training jobs are repeatedly preempted by higher-priority work but restart from scratch because the scheduler lacks recovery hooks. The associated response would be to coordinate preemption policy with checkpoint and requeue support for eligible training jobs. It does not describe the central reason for this case, which is that multi-tenant accelerator scheduling must balance fairness and business priority rather than first-come volume alone.
D:That principle is relevant to this distinct case: Researchers request four GPUs from Slurm but jobs are assigned to nodes without enough contiguous available devices. The associated response would be to configure scheduler GPU resource accounting and validated device placement constraints. It does not describe the central reason for this case, which is that multi-tenant accelerator scheduling must balance fairness and business priority rather than first-come volume alone.
Question 210
The proposed approach is to identify accessible accelerated computation as an enabler of feasible large-scale model training. In the following setting, a research group trains larger neural networks after acquiring accelerators with much higher parallel throughput. Which principle BEST supports the approach?
- Distributed learning benefits from improved networking and collective communication in addition to GPU compute.
- On-demand infrastructure reduces initial capital barriers to experimentation for intermittent workloads.
- Higher compute throughput makes previously expensive optimization workloads practical at useful scale.
- Reusable pretrained representations reduce the data and compute needed for many specialized tasks.
Explanation
Correct Answer
C:The governing principle is that higher compute throughput makes previously expensive optimization workloads practical at useful scale. It supports the measure to identify accessible accelerated computation as an enabler of feasible large-scale model training. because a research group trains larger neural networks after acquiring accelerators with much higher parallel throughput. Without the control, a workload may be described as software-only even when it cannot complete within available CPU time. The most useful acceptance evidence is: A benchmark reports training-step time and total project feasibility on the supported accelerator stack.
Incorrect Answers
A:That principle is relevant to this distinct case: A multi-node project scales after faster network fabrics reduce the time workers spend synchronizing gradients. The associated response would be to include interconnect performance among the conditions supporting modern distributed AI training progress. It does not describe the central reason for this case, which is that higher compute throughput makes previously expensive optimization workloads practical at useful scale.
B:That principle is relevant to this distinct case: A startup needs accelerator capacity only during monthly training bursts and cannot finance an idle cluster. The associated response would be to evaluate elastic cloud accelerator access as an adoption enabler while recording ongoing costs and data transfer. It does not describe the central reason for this case, which is that higher compute throughput makes previously expensive optimization workloads practical at useful scale.
D:That principle is relevant to this distinct case: An applied AI team starts from an existing pretrained vision model rather than training its representation from random weights. The associated response would be to use transfer learning appropriately and validate downstream quality with the domain-specific dataset. It does not describe the central reason for this case, which is that higher compute throughput makes previously expensive optimization workloads practical at useful scale.