NCA - AI Infrastructure and Operations NCA-AIIO Practice Test 3
Question 20
In this situation, kubernetes workers contain different GPU types but all nodes are treated as equivalent for AI job placement. The team does not use hardware feature discovery and node labels or resource constraints for workload-specific placement. Which consequence BEST characterizes the gap?
- A model needing a specific GPU feature could be sent to an incompatible node.
- A large training queue can consume resources needed for customer inference requests.
- Partially admitted distributed jobs can occupy GPUs while making no training progress.
- Priority enforcement can waste GPU hours when jobs cannot resume from checkpoint state.
Explanation
Correct Answer
A:The primary exposure is that a model needing a specific GPU feature could be sent to an incompatible node. This follows from kubernetes workers contain different GPU types but all nodes are treated as equivalent for AI job placement. where the omitted measure was to use hardware feature discovery and node labels or resource constraints for workload-specific placement. The technical reason is that schedulers need trustworthy accelerator capability metadata to place workloads on compatible devices. A good operational validation after repair would be: A scheduling test places the constrained job on a qualifying node and rejects unsupported nodes.
Incorrect Answers
B:This exposure arises in a different situation: An AI platform runs model-serving jobs and distributed training under one global namespace with no capacity boundaries. There, the corresponding safeguard is to separate workload classes with appropriate quotas, priorities and isolation rules in the scheduler. In the present case, the more direct risk is that a model needing a specific GPU feature could be sent to an incompatible node.
C:This exposure arises in a different situation: A team submits two tightly coupled distributed training pods but one starts long before resources for the other are ready. There, the corresponding safeguard is to use gang-style scheduling or coordinated job admission where supported for dependent distributed workers. In the present case, the more direct risk is that a model needing a specific GPU feature could be sent to an incompatible node.
D: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 a model needing a specific GPU feature could be sent to an incompatible node.
Question 21
The following safeguard was deployed: Select a time-series forecasting approach and test prediction error over future held-out periods. What operational observation would demonstrate its success?
- Backtesting uses chronologically separated windows and reports error across season and demand levels.
- A test set verifies citations against the actual knowledge articles and escalation of unsupported requests.
- A production-line trial records missed defects and false rejects across lighting and product variants.
- Held-out images are scored against expert annotations and a reviewer workflow remains mandatory.
Explanation
Correct Answer
A:The strongest proof is: Backtesting uses chronologically separated windows and reports error across season and demand levels. It directly tests the safeguard to select a time-series forecasting approach and test prediction error over future held-out periods. The control addresses this specific context: An energy provider forecasts electricity demand from time-series usage, calendar and weather data. It matters because forecasting predicts future numeric patterns and requires avoiding training on information not available at prediction time. Without it, data leakage can produce artificially low test error and poor live demand forecasts.
Incorrect Answers
B:This observation would validate a different control: Combine retrieval-grounded generation with source attribution and a human escalation policy. It fits the case where a customer-service team wants an LLM to draft responses using approved knowledge-base articles. It cannot prove success here; instead the required evidence is backtesting uses chronologically separated windows and reports error across season and demand levels.
C:This observation would validate a different control: Apply visual defect detection with production-representative examples and inspection feedback. It fits the case where a factory collects camera images of assembly lines to detect defective products before shipment. It cannot prove success here; instead the required evidence is backtesting uses chronologically separated windows and reports error across season and demand levels.
D: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 backtesting uses chronologically separated windows and reports error across season and demand levels.
Question 22
A team omits this measure: Estimate WAN transfer windows and evaluate approved data staging or local preprocessing to reduce movement. In this environment: A hybrid project keeps raw data on site but wants to train in the cloud using repeated full-dataset uploads. Which resulting problem is MOST likely?
- Moving datasets into an unapproved region may breach policy even if model training works.
- Placing latency-critical inference on interruptible machines may create unacceptable service disruption.
- A cloud GPU fleet may remain idle while large input datasets cross a limited WAN link.
- A low-looking hardware price may conceal ongoing operational costs and extended outages.
Explanation
Correct Answer
C:The primary exposure is that a cloud GPU fleet may remain idle while large input datasets cross a limited WAN link. This follows from a hybrid project keeps raw data on site but wants to train in the cloud using repeated full-dataset uploads. where the omitted measure was to estimate WAN transfer windows and evaluate approved data staging or local preprocessing to reduce movement. The technical reason is that hybrid AI performance depends on data locality and network movement as well as rented GPU speed. A good operational validation after repair would be: An end-to-end run measures data staging time and total cost rather than reporting only GPU kernel time.
Incorrect Answers
A:This exposure arises in a different situation: A government agency trains on restricted records that cannot leave an accredited local facility. There, the corresponding safeguard is to select an on-premises or explicitly approved sovereign environment consistent with data-residency rules. In the present case, the more direct risk is that a cloud GPU fleet may remain idle while large input datasets cross a limited WAN link.
B:This exposure arises in a different situation: An AI service has predictable daily demand but a training task can be interrupted and resumed later. There, the corresponding safeguard is to select stable serving capacity and an appropriate lower-cost interruptible training tier with checkpoint recovery. In the present case, the more direct risk is that a cloud GPU fleet may remain idle while large input datasets cross a limited WAN link.
D:This exposure arises in a different situation: A startup wants a fully owned GPU platform but lacks specialists to operate switching, cooling and firmware updates. There, the corresponding safeguard is to include operational staffing and maintenance lifecycle in the on-premises versus managed-cloud decision. In the present case, the more direct risk is that a cloud GPU fleet may remain idle while large input datasets cross a limited WAN link.
Question 23
A GPU operations plan calls for: Verify supported vGPU software version compatibility and the applicable licensing lifecycle. The reason is that an IT department deploys vGPU workloads but ignores required enterprise license and guest-driver version alignment. Which principle is MOST relevant?
- Accelerated virtual infrastructure needs supported drivers and entitlements as well as physical GPU capacity.
- Accelerator virtualization remains constrained by physical memory and supported profile limits.
- MIG creates hardware-partitioned accelerator instances on supported GPUs for eligible workloads.
- Time slicing shares execution opportunity but does not provide the same hardware isolation properties as MIG.
Explanation
Correct Answer
A:The governing principle is that accelerated virtual infrastructure needs supported drivers and entitlements as well as physical GPU capacity. It supports the measure to verify supported vGPU software version compatibility and the applicable licensing lifecycle. because an IT department deploys vGPU workloads but ignores required enterprise license and guest-driver version alignment. Without the control, workloads may lose accelerated features or support coverage even when a GPU is physically present. The most useful acceptance evidence is: A pilot VM reports valid entitlement and compatible host/guest stack during the required workload.
Incorrect Answers
B:That principle is relevant to this distinct case: A virtualization cluster admits more virtual-GPU memory assignments than can be backed by the available hardware profiles. The associated response would be to plan explicit virtual-GPU profile capacity and admission controls instead of assuming limitless overcommit. It does not describe the central reason for this case, which is that accelerated virtual infrastructure needs supported drivers and entitlements as well as physical GPU capacity.
C:That principle is relevant to this distinct case: A trusted internal platform needs to run several smaller inference services on one MIG-capable GPU with reserved resources for each workload. The associated response would be to assess GPU Multi-Instance GPU partitions on supported hardware with compatible workload sizes. It does not describe the central reason for this case, which is that accelerated virtual infrastructure needs supported drivers and entitlements as well as physical GPU capacity.
D:That principle is relevant to this distinct case: A team proposes GPU time slicing as equivalent to hardware partitioning with independent memory fault isolation. The associated response would be to distinguish time-shared GPU execution from supported MIG hardware isolation before tenant placement. It does not describe the central reason for this case, which is that accelerated virtual infrastructure needs supported drivers and entitlements as well as physical GPU capacity.
Question 24
An organization is evaluating an AI deployment: A GPU hall relies on one unprotected utility feed although the business requires continued operation through feed faults. What solution MOST directly fits the stated need?
- Design documented cable routes, bend radii and labeling for power and network infrastructure.
- Check rack access, airflow and maintenance clearances before the final deployment layout.
- Obtain qualified structural review of populated rack weight and floor-loading constraints.
- Assess electrical redundancy, UPS autonomy and generator integration against the site's availability objective.
Explanation
Correct Answer
D:The appropriate response is to assess electrical redundancy, UPS autonomy and generator integration against the site's availability objective. This is specific to the case where a GPU hall relies on one unprotected utility feed although the business requires continued operation through feed faults. The mechanism is that facility availability depends on the complete power chain, not only server dual power supplies. If ignored, a short upstream feed interruption can stop every accelerator node in the hall. Validate the choice using this targeted evidence: A qualified power-chain review maps independent paths and documented recovery actions for feed loss.
Incorrect Answers
A:This alternative addresses a different requirement: A new cluster has redundant high-speed links but cable trays cannot support the planned fiber and copper bundles. It would be appropriate there to design documented cable routes, bend radii and labeling for power and network infrastructure. But the present scenario is about a GPU hall relies on one unprotected utility feed although the business requires continued operation through feed faults. For it, the needed control is to assess electrical redundancy, UPS autonomy and generator integration against the site's availability objective.
B:This alternative addresses a different requirement: A data center orders accelerated racks with rear-door heat exchangers but leaves insufficient service clearance. It would be appropriate there to check rack access, airflow and maintenance clearances before the final deployment layout. But the present scenario is about a GPU hall relies on one unprotected utility feed although the business requires continued operation through feed faults. For it, the needed control is to assess electrical redundancy, UPS autonomy and generator integration against the site's availability objective.
C:This alternative addresses a different requirement: A facility plans high-density GPU racks on an upper floor without confirming structural load limits. It would be appropriate there to obtain qualified structural review of populated rack weight and floor-loading constraints. But the present scenario is about a GPU hall relies on one unprotected utility feed although the business requires continued operation through feed faults. For it, the needed control is to assess electrical redundancy, UPS autonomy and generator integration against the site's availability objective.
Question 25
In this situation, a model trained in PyTorch produces slow serving latency when deployed unchanged as a general-purpose Python process. The team does not evaluate TensorRT optimization and supported precision profiles for the deployed inference model and hardware. Which consequence BEST characterizes the gap?
- Teams can operate unsupported framework combinations and lose a clear path for patches or incident escalation.
- The service may miss latency and throughput targets even when model accuracy remains acceptable in offline tests.
- The preparation stage can dominate wall-clock time even when model training itself is highly accelerated.
- A customization program may lack reproducible evaluations and produce regressions after subsequent training runs.
Explanation
Correct Answer
B:The primary exposure is that the service may miss latency and throughput targets even when model accuracy remains acceptable in offline tests. This follows from a model trained in PyTorch produces slow serving latency when deployed unchanged as a general-purpose Python process. where the omitted measure was to evaluate TensorRT optimization and supported precision profiles for the deployed inference model and hardware. The technical reason is that tensorRT targets optimized inference execution through graph and kernel transformations on supported accelerators. A good operational validation after repair would be: An accuracy-checked optimized engine meets the specified p95 latency at representative request concurrency.
Incorrect Answers
A:This exposure arises in a different situation: A regulated enterprise wants tested NVIDIA AI software components with predictable support paths across GPU servers. There, the corresponding safeguard is to assess NVIDIA AI Enterprise supported software, lifecycle terms and infrastructure compatibility before adoption. In the present case, the more direct risk is that the service may miss latency and throughput targets even when model accuracy remains acceptable in offline tests.
C:This exposure arises in a different situation: An analytics department runs large tabular transformations on CPUs while its GPUs remain idle between training jobs. There, the corresponding safeguard is to assess RAPIDS GPU-accelerated data processing where the data format and operators are compatible. In the present case, the more direct risk is that the service may miss latency and throughput targets even when model accuracy remains acceptable in offline tests.
D:This exposure arises in a different situation: An LLM customization team needs a repeatable workflow for model adaptation and responsible output evaluation instead of ad hoc scripts. There, the corresponding safeguard is to evaluate NVIDIA NeMo components appropriate for model customization and evaluation within the supported stack. In the present case, the more direct risk is that the service may miss latency and throughput targets even when model accuracy remains acceptable in offline tests.
Question 26
A data-center manager declines to calculate fabric bisection bandwidth and oversubscription for the planned scale-out job mix. The operational context is: A leaf-spine AI network uses heavily oversubscribed uplinks and passes only simple single-node benchmarks. Which risk is MOST direct?
- One tenant's traffic or compromised container could disrupt shared management services.
- Link congestion may erase expected multi-node speedups while GPU computation remains efficient.
- Network links may saturate before the GPU completes its available compute work.
- Collective operations can collapse in performance as more racks become active.
Explanation
Correct Answer
D:The primary exposure is that collective operations can collapse in performance as more racks become active. This follows from a leaf-spine AI network uses heavily oversubscribed uplinks and passes only simple single-node benchmarks. where the omitted measure was to calculate fabric bisection bandwidth and oversubscription for the planned scale-out job mix. The technical reason is that distributed AI networks must be sized for realistic many-to-many communication patterns. A good operational validation after repair would be: A multi-rack benchmark shows job throughput and congestion stay within explicit scale targets.
Incorrect Answers
A:This exposure arises in a different situation: A shared GPU cluster combines tenant inference traffic and privileged management access on one unmanaged network. There, the corresponding safeguard is to segment workload and management paths with required access control and monitored boundaries. In the present case, the more direct risk is that collective operations can collapse in performance as more racks become active.
B:This exposure arises in a different situation: A multi-node transformer training job exchanges large gradients across all GPU workers after each step. There, the corresponding safeguard is to size east-west network bandwidth and collective latency for the intended distributed training topology. In the present case, the more direct risk is that collective operations can collapse in performance as more racks become active.
C:This exposure arises in a different situation: An accelerated node uses a slow network interface although its GPU data path needs high-bandwidth peer transfers. There, the corresponding safeguard is to select supported high-speed NIC or HCA capacity aligned with expected accelerator and storage traffic. In the present case, the more direct risk is that collective operations can collapse in performance as more racks become active.
Question 27
An AI infrastructure team faces this requirement: Technicians take shared GPU nodes offline during business hours without coordinating active checkpointing jobs. Which action BEST addresses the operational constraint?
- Use governed change records, dependency checks and validated rollback procedures for cluster-wide updates.
- Scope remediation to verified affected nodes and require approval for destructive maintenance actions.
- Track usable versus installed GPU hours, allocation, job backlog and maintenance losses before purchasing.
- Enforce maintenance windows with scheduler drain, workload notice and readiness verification.
Explanation
Correct Answer
D:The appropriate response is to enforce maintenance windows with scheduler drain, workload notice and readiness verification. This is specific to the case where technicians take shared GPU nodes offline during business hours without coordinating active checkpointing jobs. The mechanism is that orderly maintenance requires coordination between facilities work and the active GPU job scheduler. If ignored, unplanned repairs may destroy unsaved training state or block critical inference workloads. Validate the choice using this targeted evidence: A planned drain permits existing work to finish or checkpoint before the node enters maintenance.
Incorrect Answers
A:This alternative addresses a different requirement: A team changes cluster networking and driver versions on the same day without an approval window or recovery plan. It would be appropriate there to use governed change records, dependency checks and validated rollback procedures for cluster-wide updates. But the present scenario is about technicians take shared GPU nodes offline during business hours without coordinating active checkpointing jobs. For it, the needed control is to enforce maintenance windows with scheduler drain, workload notice and readiness verification.
B:This alternative addresses a different requirement: Engineers trigger automated node remediation from broad alerts without checking device identities or active jobs. It would be appropriate there to scope remediation to verified affected nodes and require approval for destructive maintenance actions. But the present scenario is about technicians take shared GPU nodes offline during business hours without coordinating active checkpointing jobs. For it, the needed control is to enforce maintenance windows with scheduler drain, workload notice and readiness verification.
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 technicians take shared GPU nodes offline during business hours without coordinating active checkpointing jobs. For it, the needed control is to enforce maintenance windows with scheduler drain, workload notice and readiness verification.
Question 28
The team recommends this safeguard: Evaluate a suitably sized InfiniBand fabric with documented latency, bandwidth and management requirements. It applies because a high-performance computing group trains large models requiring fast node-to-node collective operations. Which concept justifies the design?
- High-performance InfiniBand networks can support latency-sensitive collective traffic between AI nodes.
- Network routing and load distribution influence collective tail latency as much as raw line rate.
- Local GPU interconnects and external rack fabrics serve different scopes of the data path.
- Physical optical characteristics constrain achievable port speed and link reliability.
Explanation
Correct Answer
A:The governing principle is that high-performance InfiniBand networks can support latency-sensitive collective traffic between AI nodes. It supports the measure to evaluate a suitably sized InfiniBand fabric with documented latency, bandwidth and management requirements. because a high-performance computing group trains large models requiring fast node-to-node collective operations. Without the control, a fabric not designed for collectives may make distributed scaling inefficient. The most useful acceptance evidence is: Measured collective performance meets the training job's communication budget across the intended node count.
Incorrect Answers
B:That principle is relevant to this distinct case: A large network shows uneven path utilization during all-to-all traffic despite ample aggregate switch capacity. The associated response would be to evaluate supported adaptive or congestion-aware routing behavior against application traffic distribution. It does not describe the central reason for this case, which is that high-performance InfiniBand networks can support latency-sensitive collective traffic between AI nodes.
C:That principle is relevant to this distinct case: A scale-up system exchanges large tensors between GPUs inside one chassis and compares that traffic with rack networking. The associated response would be to use supported NVLink or NVSwitch paths for local GPU exchange rather than assuming every copy crosses Ethernet. It does not describe the central reason for this case, which is that high-performance InfiniBand networks can support latency-sensitive collective traffic between AI nodes.
D:That principle is relevant to this distinct case: A planned GPU cluster spans distant data halls but its optical cabling design ignores reach and transceiver compatibility. The associated response would be to select supported optics, fiber media and loss budgets for the required data-center link distances. It does not describe the central reason for this case, which is that high-performance InfiniBand networks can support latency-sensitive collective traffic between AI nodes.
Question 29
A data-center manager declines to account for inference KV-cache growth and concurrency while selecting context limits and serving memory budgets. The operational context is: A chat model serves increasingly long conversations and runs out of GPU memory despite stable model weights. Which risk is MOST direct?
- A single device cannot hold all required model states and distributed steps may become communication bound.
- Serving costs remain high while large batches can also cause unacceptable response latency.
- The service can trigger out-of-memory errors only after enough concurrent long-context sessions accumulate.
- Defaulting to full-parameter updates can exceed memory budgets without improving the target task.
Explanation
Correct Answer
C:The primary exposure is that the service can trigger out-of-memory errors only after enough concurrent long-context sessions accumulate. This follows from a chat model serves increasingly long conversations and runs out of GPU memory despite stable model weights. where the omitted measure was to account for inference KV-cache growth and concurrency while selecting context limits and serving memory budgets. The technical reason is that autoregressive inference memory includes per-session runtime state beyond the static model parameter footprint. A good operational validation after repair would be: A load test with maximum allowed context and session count stays within monitored GPU memory headroom.
Incorrect Answers
A:This exposure arises in a different situation: A foundation-model training plan exceeds the memory of a single accelerator even after supported precision reductions. There, the corresponding safeguard is to evaluate model parallelism with appropriate tensor or pipeline partitioning and high-bandwidth interconnect needs. In the present case, the more direct risk is that the service can trigger out-of-memory errors only after enough concurrent long-context sessions accumulate.
B:This exposure arises in a different situation: An inference service receives many short requests but operates each GPU one request at a time with poor utilization. There, the corresponding safeguard is to evaluate dynamic batching within the permitted queueing delay and model-specific latency budget. In the present case, the more direct risk is that the service can trigger out-of-memory errors only after enough concurrent long-context sessions accumulate.
D:This exposure arises in a different situation: A business fine-tunes a foundation model using a small private dataset and wants to reduce training resource expense. There, the corresponding safeguard is to compare parameter-efficient tuning against full-model training based on quality and available accelerator memory. In the present case, the more direct risk is that the service can trigger out-of-memory errors only after enough concurrent long-context sessions accumulate.
Question 30
Consider this NVIDIA AI infrastructure scenario: A regulated enterprise trains models locally but wants cloud inference for geographically distributed customers. Which design choice should be made FIRST?
- Estimate WAN transfer windows and evaluate approved data staging or local preprocessing to reduce movement.
- Design a hybrid model lifecycle with approved artifact replication and separate sensitive-data access rules.
- Include operational staffing and maintenance lifecycle in the on-premises versus managed-cloud decision.
- Compare elastic cloud accelerator rental against on-premises ownership using realistic utilization assumptions.
Explanation
Correct Answer
B:The appropriate response is to design a hybrid model lifecycle with approved artifact replication and separate sensitive-data access rules. This is specific to the case where a regulated enterprise trains models locally but wants cloud inference for geographically distributed customers. The mechanism is that hybrid deployments can separate training and serving but still require trustworthy provenance and data governance. If ignored, an unreviewed model artifact or customer dataset may cross a security boundary during deployment. Validate the choice using this targeted evidence: A controlled promotion records model identity and proves inference receives only permitted data in each region.
Incorrect Answers
A:This alternative addresses a different requirement: A hybrid project keeps raw data on site but wants to train in the cloud using repeated full-dataset uploads. It would be appropriate there to estimate WAN transfer windows and evaluate approved data staging or local preprocessing to reduce movement. But the present scenario is about a regulated enterprise trains models locally but wants cloud inference for geographically distributed customers. For it, the needed control is to design a hybrid model lifecycle with approved artifact replication and separate sensitive-data access rules.
C:This alternative addresses a different requirement: A startup wants a fully owned GPU platform but lacks specialists to operate switching, cooling and firmware updates. It would be appropriate there to include operational staffing and maintenance lifecycle in the on-premises versus managed-cloud decision. But the present scenario is about a regulated enterprise trains models locally but wants cloud inference for geographically distributed customers. For it, the needed control is to design a hybrid model lifecycle with approved artifact replication and separate sensitive-data access rules.
D:This alternative addresses a different requirement: A small research lab runs GPU training only six days per quarter and is deciding whether to purchase a large fleet. It would be appropriate there to compare elastic cloud accelerator rental against on-premises ownership using realistic utilization assumptions. But the present scenario is about a regulated enterprise trains models locally but wants cloud inference for geographically distributed customers. For it, the needed control is to design a hybrid model lifecycle with approved artifact replication and separate sensitive-data access rules.