NCA - AI Infrastructure and Operations NCA-AIIO Practice Test 18
Question 170
The team recommends this safeguard: Correlate accelerator temperature, power limit and clock throttling telemetry with ambient conditions. It applies because an operations team sees lower training throughput on warm afternoons and suspects application regressions. Which concept justifies the design?
- High-density liquid cooling requires compatible facility infrastructure beyond server procurement.
- Data-center maintenance planning must account for the coupling between compute load and cooling availability.
- Local thermal conditions determine equipment behavior more directly than a remote room thermostat.
- Thermal or power constraints can reduce effective GPU clocks without changing model code.
Explanation
Correct Answer
D:The governing principle is that thermal or power constraints can reduce effective GPU clocks without changing model code. It supports the measure to correlate accelerator temperature, power limit and clock throttling telemetry with ambient conditions. because an operations team sees lower training throughput on warm afternoons and suspects application regressions. Without the control, engineers may misdiagnose heat-related throttling as a defective training framework. The most useful acceptance evidence is: Monitoring aligns clock reductions with temperature or power-limit events and a cooling fix restores speed.
Incorrect Answers
A:That principle is relevant to this distinct case: A team wants to deploy liquid-cooled GPU systems in a building that has no validated coolant distribution network. The associated response would be to check facility liquid cooling availability, heat exchangers, leak detection and vendor flow requirements. It does not describe the central reason for this case, which is that thermal or power constraints can reduce effective GPU clocks without changing model code.
B:That principle is relevant to this distinct case: Facilities schedules maintenance to an air-handling unit while operations keeps every GPU rack at full demand. The associated response would be to coordinate workload reduction and protected shutdown windows against available cooling redundancy. It does not describe the central reason for this case, which is that thermal or power constraints can reduce effective GPU clocks without changing model code.
C:That principle is relevant to this distinct case: A rack runs multiple powerful accelerators but operators monitor only average room temperature, not intake air. The associated response would be to collect rack inlet and component thermal measurements during representative peak workloads. It does not describe the central reason for this case, which is that thermal or power constraints can reduce effective GPU clocks without changing model code.
Question 171
A GPU operations plan calls for: Optimize for sustained training throughput, memory capacity and batch processing rather than single-request latency. The reason is that a research team trains a large image classifier nightly with lengthy epochs and no strict per-request response deadline. Which principle is MOST relevant?
- Batching can improve throughput but adds queueing latency and must respect serving requirements.
- Training recovery needs recoverable optimizer and model state rather than only a copy of the final weights.
- Training repeatedly computes gradients and updates parameters, placing different demands on hardware than inference.
- Synchronized data-parallel training exchanges gradients, so collective efficiency influences scaling gains.
Explanation
Correct Answer
C:The governing principle is that training repeatedly computes gradients and updates parameters, placing different demands on hardware than inference. It supports the measure to optimize for sustained training throughput, memory capacity and batch processing rather than single-request latency. because a research team trains a large image classifier nightly with lengthy epochs and no strict per-request response deadline. Without the control, selecting only the lowest-latency serving profile can waste training capacity or exhaust memory on full epochs. The most useful acceptance evidence is: Representative epochs finish within the time target while convergence and checkpoint integrity remain acceptable.
Incorrect Answers
A:That principle is relevant to this distinct case: An inference service receives many short requests but operates each GPU one request at a time with poor utilization. The associated response would be to evaluate dynamic batching within the permitted queueing delay and model-specific latency budget. It does not describe the central reason for this case, which is that training repeatedly computes gradients and updates parameters, placing different demands on hardware than inference.
B:That principle is relevant to this distinct case: A multi-hour distributed model training run loses all progress when one compute node is recycled for maintenance. The associated response would be to write versioned training checkpoints to resilient storage and test restart from a known consistent state. It does not describe the central reason for this case, which is that training repeatedly computes gradients and updates parameters, placing different demands on hardware than inference.
D:That principle is relevant to this distinct case: A team scales training to many GPU nodes but iteration time rises because gradient exchanges dominate compute. The associated response would be to profile collective communication and adjust topology, batch strategy or network capacity for data-parallel training. It does not describe the central reason for this case, which is that training repeatedly computes gradients and updates parameters, placing different demands on hardware than inference.
Question 172
A data-center manager declines to track usable versus installed GPU hours, allocation, job backlog and maintenance losses before purchasing. The operational context is: A research group requests another rack of GPUs while existing capacity is partly unavailable or poorly scheduled. Which risk is MOST direct?
- A failed update can disable many training jobs with no tested route back to a known working state.
- Unreviewed version drift can create device instability or unexplained workload incompatibilities.
- Unplanned repairs may destroy unsaved training state or block critical inference workloads.
- An expansion purchase can be unnecessary while scheduling or repair problems remain unresolved.
Explanation
Correct Answer
D:The primary exposure is that an expansion purchase can be unnecessary while scheduling or repair problems remain unresolved. This follows from a research group requests another rack of GPUs while existing capacity is partly unavailable or poorly scheduled. where the omitted measure was to track usable versus installed GPU hours, allocation, job backlog and maintenance losses before purchasing. The technical reason is that capacity planning must distinguish installed accelerator count from reliable available processing capacity. A good operational validation after repair would be: A utilization report reconciles GPU availability, queue time and recoverable outage loss across the fleet.
Incorrect Answers
A:This exposure arises in a different situation: A team changes cluster networking and driver versions on the same day without an approval window or recovery plan. There, the corresponding safeguard is to use governed change records, dependency checks and validated rollback procedures for cluster-wide updates. In the present case, the more direct risk is that an expansion purchase can be unnecessary while scheduling or repair problems remain unresolved.
B:This exposure arises in a different situation: A service engineer upgrades GPU firmware on one node without recording it, leading to mismatched cluster behavior. There, the corresponding safeguard is to use a governed firmware baseline and controlled upgrade or rollback with node-level inventory comparison. In the present case, the more direct risk is that an expansion purchase can be unnecessary while scheduling or repair problems remain unresolved.
C:This exposure arises in a different situation: Technicians take shared GPU nodes offline during business hours without coordinating active checkpointing jobs. There, the corresponding safeguard is to enforce maintenance windows with scheduler drain, workload notice and readiness verification. In the present case, the more direct risk is that an expansion purchase can be unnecessary while scheduling or repair problems remain unresolved.
Question 173
Consider this NVIDIA AI infrastructure scenario: A leaf-spine AI network uses heavily oversubscribed uplinks and passes only simple single-node benchmarks. Which design choice should be made FIRST?
- Estimate concurrent network needs and evaluate storage-traffic isolation or quality-of-service policies.
- Calculate fabric bisection bandwidth and oversubscription for the planned scale-out job mix.
- Evaluate supported GPUDirect RDMA and the required GPU, NIC, driver and topology compatibility.
- Size east-west network bandwidth and collective latency for the intended distributed training topology.
Explanation
Correct Answer
B:The appropriate response is to calculate fabric bisection bandwidth and oversubscription for the planned scale-out job mix. This is specific to the case where a leaf-spine AI network uses heavily oversubscribed uplinks and passes only simple single-node benchmarks. The mechanism is that distributed AI networks must be sized for realistic many-to-many communication patterns. If ignored, collective operations can collapse in performance as more racks become active. Validate the choice using this targeted evidence: A multi-rack benchmark shows job throughput and congestion stay within explicit scale targets.
Incorrect Answers
A:This alternative addresses a different requirement: Cluster operators run training dataset reads and checkpoint writes through the same fabric as synchronous gradient traffic. It would be appropriate there to estimate concurrent network needs and evaluate storage-traffic isolation or quality-of-service policies. But the present scenario is about a leaf-spine AI network uses heavily oversubscribed uplinks and passes only simple single-node benchmarks. For it, the needed control is to calculate fabric bisection bandwidth and oversubscription for the planned scale-out job mix.
C:This alternative addresses a different requirement: A research team expects GPUs to access distributed data without unnecessary staging through host CPU memory. It would be appropriate there to evaluate supported GPUDirect RDMA and the required GPU, NIC, driver and topology compatibility. But the present scenario is about a leaf-spine AI network uses heavily oversubscribed uplinks and passes only simple single-node benchmarks. For it, the needed control is to calculate fabric bisection bandwidth and oversubscription for the planned scale-out job mix.
D:This alternative addresses a different requirement: A multi-node transformer training job exchanges large gradients across all GPU workers after each step. It would be appropriate there to size east-west network bandwidth and collective latency for the intended distributed training topology. But the present scenario is about a leaf-spine AI network uses heavily oversubscribed uplinks and passes only simple single-node benchmarks. For it, the needed control is to calculate fabric bisection bandwidth and oversubscription for the planned scale-out job mix.
Question 174
Consider this NVIDIA AI infrastructure scenario: A company says its production fraud-scoring API is still learning from each request although model weights never change. Which design choice should be made FIRST?
- Separate trained-model inference from explicit retraining or online-learning update processes.
- Classify deep learning as a subset of machine learning using layered neural representations.
- Distinguish generative modeling from discriminative classification when selecting model evaluation criteria.
- Distinguish fixed rule-based automation from models that learn patterns from examples during training.
Explanation
Correct Answer
A:The appropriate response is to separate trained-model inference from explicit retraining or online-learning update processes. This is specific to the case where a company says its production fraud-scoring API is still learning from each request although model weights never change. The mechanism is that inference normally applies learned parameters; serving requests alone does not imply those parameters update. If ignored, operators can overlook model drift while assuming a deployed checkpoint adapts automatically. Validate the choice using this targeted evidence: A model-version audit shows fixed parameters during inference and a separate governed update process.
Incorrect Answers
B:This alternative addresses a different requirement: An enterprise compares a decision tree with a deep multilayer vision network and claims both use identical learning architectures. It would be appropriate there to classify deep learning as a subset of machine learning using layered neural representations. But the present scenario is about a company says its production fraud-scoring API is still learning from each request although model weights never change. For it, the needed control is to separate trained-model inference from explicit retraining or online-learning update processes.
C:This alternative addresses a different requirement: A text model generates new summaries while a separate classifier assigns each message to a fixed department. It would be appropriate there to distinguish generative modeling from discriminative classification when selecting model evaluation criteria. But the present scenario is about a company says its production fraud-scoring API is still learning from each request although model weights never change. For it, the needed control is to separate trained-model inference from explicit retraining or online-learning update processes.
D:This alternative addresses a different requirement: A project hard-codes hundreds of manually chosen rules to classify support tickets, yet the team describes it as a trained ML model. It would be appropriate there to distinguish fixed rule-based automation from models that learn patterns from examples during training. But the present scenario is about a company says its production fraud-scoring API is still learning from each request although model weights never change. For it, the needed control is to separate trained-model inference from explicit retraining or online-learning update processes.
Question 175
The following safeguard was deployed: Segment workload and management paths with required access control and monitored boundaries. What operational observation would demonstrate its success?
- A multi-rack benchmark shows job throughput and congestion stay within explicit scale targets.
- An all-reduce test measures steady collective latency and bandwidth under representative scale.
- A data transfer benchmark shows per-node port throughput within the planned utilization envelope.
- A segmentation test blocks unauthorized tenant-to-management traffic while approved operations continue.
Explanation
Correct Answer
D:The strongest proof is: A segmentation test blocks unauthorized tenant-to-management traffic while approved operations continue. It directly tests the safeguard to segment workload and management paths with required access control and monitored boundaries. The control addresses this specific context: A shared GPU cluster combines tenant inference traffic and privileged management access on one unmanaged network. It matters because network isolation limits unauthorized access and protects operational control from workload contention. Without it, one tenant's traffic or compromised container could disrupt shared management services.
Incorrect Answers
A:This observation would validate a different control: Calculate fabric bisection bandwidth and oversubscription for the planned scale-out job mix. It fits the case where a leaf-spine AI network uses heavily oversubscribed uplinks and passes only simple single-node benchmarks. It cannot prove success here; instead the required evidence is a segmentation test blocks unauthorized tenant-to-management traffic while approved operations continue.
B:This observation would validate a different control: Size east-west network bandwidth and collective latency for the intended distributed training topology. It fits the case where a multi-node transformer training job exchanges large gradients across all GPU workers after each step. It cannot prove success here; instead the required evidence is a segmentation test blocks unauthorized tenant-to-management traffic while approved operations continue.
C:This observation would validate a different control: Select supported high-speed NIC or HCA capacity aligned with expected accelerator and storage traffic. It fits the case where an accelerated node uses a slow network interface although its GPU data path needs high-bandwidth peer transfers. It cannot prove success here; instead the required evidence is a segmentation test blocks unauthorized tenant-to-management traffic while approved operations continue.
Question 176
The proposed approach is to plan checkpoint intervals, scheduler restart rules and failure-domain containment for distributed jobs. In the following setting, a GPU cluster loses progress whenever a worker disappears during a large multi-node training epoch. Which principle BEST supports the approach?
- Scale-up and scale-out use different interconnect domains and communication constraints.
- Adding more distributed nodes expands the operational failure surface as well as compute capacity.
- Data parallelism replicates the model and partitions data while coordinating parameter updates.
- Distributed accelerator capacity cannot scale efficiently when network bisection bandwidth is insufficient.
Explanation
Correct Answer
B:The governing principle is that adding more distributed nodes expands the operational failure surface as well as compute capacity. It supports the measure to plan checkpoint intervals, scheduler restart rules and failure-domain containment for distributed jobs. because a GPU cluster loses progress whenever a worker disappears during a large multi-node training epoch. Without the control, a transient node outage can waste all work since the last recoverable coordinated checkpoint. The most useful acceptance evidence is: An authorized node-failure exercise resumes from a valid checkpoint without corrupt model state.
Incorrect Answers
A: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 adding more distributed nodes expands the operational failure surface as well as compute capacity.
C:That principle is relevant to this distinct case: A model fits on one GPU but an organization wants to speed up training using more GPUs on the same dataset. The associated response would be to evaluate data-parallel workers with synchronized gradients and a suitable global batch strategy. It does not describe the central reason for this case, which is that adding more distributed nodes expands the operational failure surface as well as compute capacity.
D:That principle is relevant to this distinct case: A data center scales from four to sixty-four training nodes without upgrading per-node network ports. The associated response would be to model aggregate east-west traffic and network oversubscription at the intended cluster scale. It does not describe the central reason for this case, which is that adding more distributed nodes expands the operational failure surface as well as compute capacity.
Question 177
A data-center manager declines to gate retraining output on dataset validation, safety metrics and approved promotion criteria. The operational context is: An organization automatically retrains a model nightly and publishes weights without independent acceptance rules. Which risk is MOST direct?
- An incident team may deploy untested weights or lose the ability to reproduce a past prediction.
- Repeated test-set tuning can exaggerate real-world quality and lead to unsafe deployment decisions.
- A faulty input batch can cause an inferior model to replace a stable production version.
- An accurate checkpoint can become unavailable or break client requests after release.
Explanation
Correct Answer
C:The primary exposure is that a faulty input batch can cause an inferior model to replace a stable production version. This follows from an organization automatically retrains a model nightly and publishes weights without independent acceptance rules. where the omitted measure was to gate retraining output on dataset validation, safety metrics and approved promotion criteria. The technical reason is that automated retraining is a governed change-management process, not an unconditional parameter refresh. A good operational validation after repair would be: A deliberately degraded candidate is rejected while the previously approved model keeps serving.
Incorrect Answers
A:This exposure arises in a different situation: Operations cannot determine which trained model is running in each deployment because artifacts are overwritten under one label. There, the corresponding safeguard is to register immutable model versions with provenance, evaluation results and approved release transitions. In the present case, the more direct risk is that a faulty input batch can cause an inferior model to replace a stable production version.
B: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 a faulty input batch can cause an inferior model to replace a stable production version.
D:This exposure arises in a different situation: A model-serving change reaches production without checking its API schema, resource limits or startup health. There, the corresponding safeguard is to use staged deployment validation for model artifact, endpoint compatibility, GPU resources and readiness. In the present case, the more direct risk is that a faulty input batch can cause an inferior model to replace a stable production version.
Question 178
The following safeguard was deployed: Preload supported model artifacts or maintain warm capacity based on startup latency and expected demand. What operational observation would demonstrate its success?
- A scaling test records useful samples per second and all-reduce time at each cluster size.
- The sharded training run completes one correct optimization step with measured communication overhead.
- A sudden traffic test meets service latency objectives while startup and readiness times are recorded.
- An evaluation compares task accuracy and peak memory for an approved adapter versus a full-tune baseline.
Explanation
Correct Answer
C:The strongest proof is: A sudden traffic test meets service latency objectives while startup and readiness times are recorded. It directly tests the safeguard to preload supported model artifacts or maintain warm capacity based on startup latency and expected demand. The control addresses this specific context: An autoscaled text-generation endpoint starts new model replicas only after a demand spike and sees long cold starts. It matters because inference scale-out must include model loading and initialization time, not only GPU allocation time. Without it, users may experience a prolonged latency spike while newly scheduled replicas initialize model weights.
Incorrect Answers
A:This observation would validate a different control: Profile collective communication and adjust topology, batch strategy or network capacity for data-parallel training. It fits the case where a team scales training to many GPU nodes but iteration time rises because gradient exchanges dominate compute. It cannot prove success here; instead the required evidence is a sudden traffic test meets service latency objectives while startup and readiness times are recorded.
B:This observation would validate a different control: Evaluate model parallelism with appropriate tensor or pipeline partitioning and high-bandwidth interconnect needs. It fits the case where a foundation-model training plan exceeds the memory of a single accelerator even after supported precision reductions. It cannot prove success here; instead the required evidence is a sudden traffic test meets service latency objectives while startup and readiness times are recorded.
D:This observation would validate a different control: Compare parameter-efficient tuning against full-model training based on quality and available accelerator memory. It fits the case where a business fine-tunes a foundation model using a small private dataset and wants to reduce training resource expense. It cannot prove success here; instead the required evidence is a sudden traffic test meets service latency objectives while startup and readiness times are recorded.
Question 179
In this situation, a cloud provider standardizes on Ethernet switching but wants accelerated AI traffic over the existing operational model. The team does not assess a tuned high-speed Ethernet or RoCE solution rather than assuming ordinary LAN defaults are enough. Which consequence BEST characterizes the gap?
- A generic access-network design may fail to support predictable training traffic behavior.
- Unsupported optics or excessive signal loss may cause frequent link flaps at peak utilization.
- A fabric not designed for collectives may make distributed scaling inefficient.
- Routing local tensor transfers through a slower external path can waste accelerator time.
Explanation
Correct Answer
A:The primary exposure is that a generic access-network design may fail to support predictable training traffic behavior. This follows from a cloud provider standardizes on Ethernet switching but wants accelerated AI traffic over the existing operational model. where the omitted measure was to assess a tuned high-speed Ethernet or RoCE solution rather than assuming ordinary LAN defaults are enough. The technical reason is that modern Ethernet fabrics can support suitable AI networking when loss and congestion are managed. A good operational validation after repair would be: An approved Ethernet pilot shows stable throughput and congestion metrics under all-to-all traffic.
Incorrect Answers
B:This exposure arises in a different situation: A planned GPU cluster spans distant data halls but its optical cabling design ignores reach and transceiver compatibility. There, the corresponding safeguard is to select supported optics, fiber media and loss budgets for the required data-center link distances. In the present case, the more direct risk is that a generic access-network design may fail to support predictable training traffic behavior.
C:This exposure arises in a different situation: A high-performance computing group trains large models requiring fast node-to-node collective operations. There, the corresponding safeguard is to evaluate a suitably sized InfiniBand fabric with documented latency, bandwidth and management requirements. In the present case, the more direct risk is that a generic access-network design may fail to support predictable training traffic behavior.
D:This exposure arises in a different situation: A scale-up system exchanges large tensors between GPUs inside one chassis and compares that traffic with rack networking. There, the corresponding safeguard is to use supported NVLink or NVSwitch paths for local GPU exchange rather than assuming every copy crosses Ethernet. In the present case, the more direct risk is that a generic access-network design may fail to support predictable training traffic behavior.
Question 180
The selected AI infrastructure safeguard is not implemented: Distinguish generative modeling from discriminative classification when selecting model evaluation criteria. The observed situation is: A text model generates new summaries while a separate classifier assigns each message to a fixed department. What exposure remains?
- Testing only category accuracy can miss hallucinated or unsafe content from the generative model.
- The project may use unsuitable classification metrics and overlook unsafe exploration behavior.
- Operators can overlook model drift while assuming a deployed checkpoint adapts automatically.
- The project may promise improvements from additional data even though its rule engine never learns.
Explanation
Correct Answer
A:The primary exposure is that testing only category accuracy can miss hallucinated or unsafe content from the generative model. This follows from a text model generates new summaries while a separate classifier assigns each message to a fixed department. where the omitted measure was to distinguish generative modeling from discriminative classification when selecting model evaluation criteria. The technical reason is that generative systems create content or sequences while classifiers predict categories for observed inputs. A good operational validation after repair would be: The evaluation suite scores factual support of summaries separately from the department classification accuracy.
Incorrect Answers
B:This exposure arises in a different situation: A robot learns control actions by interacting with a simulated environment and receiving delayed reward signals. There, the corresponding safeguard is to recognize reinforcement learning and define rewards, actions and safe simulation boundaries. In the present case, the more direct risk is that testing only category accuracy can miss hallucinated or unsafe content from the generative model.
C:This exposure arises in a different situation: A company says its production fraud-scoring API is still learning from each request although model weights never change. There, the corresponding safeguard is to separate trained-model inference from explicit retraining or online-learning update processes. In the present case, the more direct risk is that testing only category accuracy can miss hallucinated or unsafe content from the generative model.
D:This exposure arises in a different situation: A project hard-codes hundreds of manually chosen rules to classify support tickets, yet the team describes it as a trained ML model. There, the corresponding safeguard is to distinguish fixed rule-based automation from models that learn patterns from examples during training. In the present case, the more direct risk is that testing only category accuracy can miss hallucinated or unsafe content from the generative model.