NCA - AI Infrastructure and Operations NCA-AIIO Practice Test 27
Question 260
A data-center engineer must resolve the following situation: A multi-node transformer training job exchanges large gradients across all GPU workers after each step. What is the MOST appropriate response?
- Size east-west network bandwidth and collective latency for the intended distributed training topology.
- Segment workload and management paths with required access control and monitored boundaries.
- Select supported high-speed NIC or HCA capacity aligned with expected accelerator and storage traffic.
- Estimate concurrent network needs and evaluate storage-traffic isolation or quality-of-service policies.
Explanation
Correct Answer
A:The appropriate response is to size east-west network bandwidth and collective latency for the intended distributed training topology. This is specific to the case where a multi-node transformer training job exchanges large gradients across all GPU workers after each step. The mechanism is that training network requirements are often driven by accelerator-to-accelerator synchronization traffic. If ignored, link congestion may erase expected multi-node speedups while GPU computation remains efficient. Validate the choice using this targeted evidence: An all-reduce test measures steady collective latency and bandwidth under representative scale.
Incorrect Answers
B:This alternative addresses a different requirement: A shared GPU cluster combines tenant inference traffic and privileged management access on one unmanaged network. It would be appropriate there to segment workload and management paths with required access control and monitored boundaries. But the present scenario is about a multi-node transformer training job exchanges large gradients across all GPU workers after each step. For it, the needed control is to size east-west network bandwidth and collective latency for the intended distributed training topology.
C:This alternative addresses a different requirement: An accelerated node uses a slow network interface although its GPU data path needs high-bandwidth peer transfers. It would be appropriate there to select supported high-speed NIC or HCA capacity aligned with expected accelerator and storage traffic. But the present scenario is about a multi-node transformer training job exchanges large gradients across all GPU workers after each step. For it, the needed control is to size east-west network bandwidth and collective latency for the intended distributed training topology.
D: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 multi-node transformer training job exchanges large gradients across all GPU workers after each step. For it, the needed control is to size east-west network bandwidth and collective latency for the intended distributed training topology.
Question 261
The design team faces this case: A cloud-based training project moves large model checkpoints between regions and ignores transfer charges. It adopts: Include data egress, cross-region transfer and retained storage costs in the deployment decision. What evidence BEST validates the intended outcome?
- The cost estimate reconciles measured storage and transfer usage with provider billing categories.
- An end-to-end run measures data staging time and total cost rather than reporting only GPU kernel time.
- The data-flow assessment confirms storage, access and compute remain inside the permitted control boundary.
- The proposed operating model assigns trained owners and budgets recurring support and facility work.
Explanation
Correct Answer
A:The strongest proof is: The cost estimate reconciles measured storage and transfer usage with provider billing categories. It directly tests the safeguard to include data egress, cross-region transfer and retained storage costs in the deployment decision. The control addresses this specific context: A cloud-based training project moves large model checkpoints between regions and ignores transfer charges. It matters because infrastructure cost modeling must count data movement and persistence, not only accelerator rental time. Without it, a seemingly inexpensive cloud training job may exceed budget due to repeated checkpoint transfers.
Incorrect Answers
B:This observation would validate a different control: Estimate WAN transfer windows and evaluate approved data staging or local preprocessing to reduce movement. It fits the case where a hybrid project keeps raw data on site but wants to train in the cloud using repeated full-dataset uploads. It cannot prove success here; instead the required evidence is the cost estimate reconciles measured storage and transfer usage with provider billing categories.
C:This observation would validate a different control: Select an on-premises or explicitly approved sovereign environment consistent with data-residency rules. It fits the case where a government agency trains on restricted records that cannot leave an accredited local facility. It cannot prove success here; instead the required evidence is the cost estimate reconciles measured storage and transfer usage with provider billing categories.
D: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 the cost estimate reconciles measured storage and transfer usage with provider billing categories.
Question 262
A technology review accepts this control: Track usable versus installed GPU hours, allocation, job backlog and maintenance losses before purchasing. Which observed result provides the STRONGEST evidence of effectiveness?
- A planned drain permits existing work to finish or checkpoint before the node enters maintenance.
- The fleet dashboard identifies the affected rack and correlates its degraded nodes with failed job records.
- A controlled sensor alarm reports chassis identity, severity and accountable response owner.
- A utilization report reconciles GPU availability, queue time and recoverable outage loss across the fleet.
Explanation
Correct Answer
D:The strongest proof is: A utilization report reconciles GPU availability, queue time and recoverable outage loss across the fleet. It directly tests the safeguard to track usable versus installed GPU hours, allocation, job backlog and maintenance losses before purchasing. The control addresses this specific context: A research group requests another rack of GPUs while existing capacity is partly unavailable or poorly scheduled. It matters because capacity planning must distinguish installed accelerator count from reliable available processing capacity. Without it, an expansion purchase can be unnecessary while scheduling or repair problems remain unresolved.
Incorrect Answers
A:This observation would validate a different control: Enforce maintenance windows with scheduler drain, workload notice and readiness verification. It fits the case where technicians take shared GPU nodes offline during business hours without coordinating active checkpointing jobs. It cannot prove success here; instead the required evidence is a utilization report reconciles GPU availability, queue time and recoverable outage loss across the fleet.
B:This observation would validate a different control: Build a central fleet dashboard joining node inventory, job states, health and GPU-capacity measures. It fits the case where an operations team tracks training-job completion but has no aggregate view of failing nodes or rack utilization. It cannot prove success here; instead the required evidence is a utilization report reconciles GPU availability, queue time and recoverable outage loss across the fleet.
C:This observation would validate a different control: Integrate supported power-supply and chassis fault signals with prioritized alert routing. It fits the case where data-center management sees correct fan speeds but receives no alert when an accelerator power module reports a fault. It cannot prove success here; instead the required evidence is a utilization report reconciles GPU availability, queue time and recoverable outage loss across the fleet.
Question 263
Consider this NVIDIA AI infrastructure scenario: A trusted internal platform needs to run several smaller inference services on one MIG-capable GPU with reserved resources for each workload. Which design choice should be made FIRST?
- Evaluate supported NVIDIA vGPU capabilities, profiles and licensing against the application requirements.
- Distinguish time-shared GPU execution from supported MIG hardware isolation before tenant placement.
- Enforce hypervisor, guest-driver and tenant resource permissions for virtualized accelerated workloads.
- Assess GPU Multi-Instance GPU partitions on supported hardware with compatible workload sizes.
Explanation
Correct Answer
D:The appropriate response is to assess GPU Multi-Instance GPU partitions on supported hardware with compatible workload sizes. This is specific to the case where a trusted internal platform needs to run several smaller inference services on one MIG-capable GPU with reserved resources for each workload. The mechanism is that mIG creates hardware-partitioned accelerator instances on supported GPUs for eligible workloads. If ignored, without appropriate partitioning, one service may consume memory needed by other workloads. Validate the choice using this targeted evidence: Two supported partitioned workloads run within their assigned memory and compute resources without interference.
Incorrect Answers
A:This alternative addresses a different requirement: A virtual-desktop deployment requires accelerated graphics for several isolated users on shared server GPUs. It would be appropriate there to evaluate supported NVIDIA vGPU capabilities, profiles and licensing against the application requirements. But the present scenario is about a trusted internal platform needs to run several smaller inference services on one MIG-capable GPU with reserved resources for each workload. For it, the needed control is to assess GPU Multi-Instance GPU partitions on supported hardware with compatible workload sizes.
B:This alternative addresses a different requirement: A team proposes GPU time slicing as equivalent to hardware partitioning with independent memory fault isolation. It would be appropriate there to distinguish time-shared GPU execution from supported MIG hardware isolation before tenant placement. But the present scenario is about a trusted internal platform needs to run several smaller inference services on one MIG-capable GPU with reserved resources for each workload. For it, the needed control is to assess GPU Multi-Instance GPU partitions on supported hardware with compatible workload sizes.
C:This alternative addresses a different requirement: An organization consolidates multiple AI tenants but has no process to constrain access to virtual GPU devices and images. It would be appropriate there to enforce hypervisor, guest-driver and tenant resource permissions for virtualized accelerated workloads. But the present scenario is about a trusted internal platform needs to run several smaller inference services on one MIG-capable GPU with reserved resources for each workload. For it, the needed control is to assess GPU Multi-Instance GPU partitions on supported hardware with compatible workload sizes.
Question 264
A data-center engineer must resolve the following situation: A cloud-based training project moves large model checkpoints between regions and ignores transfer charges. What is the MOST appropriate response?
- Compare elastic cloud accelerator rental against on-premises ownership using realistic utilization assumptions.
- Estimate WAN transfer windows and evaluate approved data staging or local preprocessing to reduce movement.
- Include data egress, cross-region transfer and retained storage costs in the deployment decision.
- Select an on-premises or explicitly approved sovereign environment consistent with data-residency rules.
Explanation
Correct Answer
C:The appropriate response is to include data egress, cross-region transfer and retained storage costs in the deployment decision. This is specific to the case where a cloud-based training project moves large model checkpoints between regions and ignores transfer charges. The mechanism is that infrastructure cost modeling must count data movement and persistence, not only accelerator rental time. If ignored, a seemingly inexpensive cloud training job may exceed budget due to repeated checkpoint transfers. Validate the choice using this targeted evidence: The cost estimate reconciles measured storage and transfer usage with provider billing categories.
Incorrect Answers
A: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 cloud-based training project moves large model checkpoints between regions and ignores transfer charges. For it, the needed control is to include data egress, cross-region transfer and retained storage costs in the deployment decision.
B: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 cloud-based training project moves large model checkpoints between regions and ignores transfer charges. For it, the needed control is to include data egress, cross-region transfer and retained storage costs in the deployment decision.
D:This alternative addresses a different requirement: A government agency trains on restricted records that cannot leave an accredited local facility. It would be appropriate there to select an on-premises or explicitly approved sovereign environment consistent with data-residency rules. But the present scenario is about a cloud-based training project moves large model checkpoints between regions and ignores transfer charges. For it, the needed control is to include data egress, cross-region transfer and retained storage costs in the deployment decision.
Question 265
Consider this NVIDIA AI infrastructure scenario: An autonomous vehicle company trains a vision policy and needs rare-road-condition examples without public-road risk. Which design choice should be made FIRST?
- Apply visual defect detection with production-representative examples and inspection feedback.
- Use high-fidelity simulation or synthetic data generation with realism and generalization testing.
- Evaluate computer-vision segmentation with clinically governed validation and human oversight.
- Consider a recommendation model trained on consented interaction data and evaluate ranking quality.
Explanation
Correct Answer
B:The appropriate response is to use high-fidelity simulation or synthetic data generation with realism and generalization testing. This is specific to the case where an autonomous vehicle company trains a vision policy and needs rare-road-condition examples without public-road risk. The mechanism is that simulation can expose AI systems to hazardous edge cases under controlled repeatable conditions. If ignored, a model may look robust in synthetic scenes but fail on real-world visual distributions. Validate the choice using this targeted evidence: The validation plan measures scenario coverage and compares simulator and real-world held-out performance.
Incorrect Answers
A:This alternative addresses a different requirement: A factory collects camera images of assembly lines to detect defective products before shipment. It would be appropriate there to apply visual defect detection with production-representative examples and inspection feedback. But the present scenario is about an autonomous vehicle company trains a vision policy and needs rare-road-condition examples without public-road risk. For it, the needed control is to use high-fidelity simulation or synthetic data generation with realism and generalization testing.
C:This alternative addresses a different requirement: A radiology team needs to highlight suspicious structures in diagnostic images for qualified human review. It would be appropriate there to evaluate computer-vision segmentation with clinically governed validation and human oversight. But the present scenario is about an autonomous vehicle company trains a vision policy and needs rare-road-condition examples without public-road risk. For it, the needed control is to use high-fidelity simulation or synthetic data generation with realism and generalization testing.
D:This alternative addresses a different requirement: A retailer wants to recommend products based on customers' prior interactions and item characteristics. It would be appropriate there to consider a recommendation model trained on consented interaction data and evaluate ranking quality. But the present scenario is about an autonomous vehicle company trains a vision policy and needs rare-road-condition examples without public-road risk. For it, the needed control is to use high-fidelity simulation or synthetic data generation with realism and generalization testing.
Question 266
The design team faces this case: A cluster sees intermittent network errors but its DPU ports and offload services are missing from health dashboards. It adopts: Monitor supported BlueField interfaces, service state and relevant packet or error counters. What evidence BEST validates the intended outcome?
- An access test demonstrates the host cannot alter protected DPU management settings without authorization.
- An injected supported offload fault is localized to the DPU service or interface in cluster telemetry.
- A controlled comparison shows lower host CPU demand without reducing storage or network correctness.
- Inventory and maintenance records show compatible host and DPU versions and a tested recovery approach.
Explanation
Correct Answer
B:The strongest proof is: An injected supported offload fault is localized to the DPU service or interface in cluster telemetry. It directly tests the safeguard to monitor supported BlueField interfaces, service state and relevant packet or error counters. The control addresses this specific context: A cluster sees intermittent network errors but its DPU ports and offload services are missing from health dashboards. It matters because dPU deployment adds operational elements that need observability alongside GPU and host status. Without it, invisible offload failures may be mistaken for model or GPU faults and prolong repair.
Incorrect Answers
A:This observation would validate a different control: Evaluate BlueField DPU-supported trusted infrastructure processing and its administrative trust boundaries. It fits the case where a security architect wants infrastructure service isolation from host applications that may be compromised. It cannot prove success here; instead the required evidence is an injected supported offload fault is localized to the DPU service or interface in cluster telemetry.
C:This observation would validate a different control: Assess supported DPU offload for suitable networking and storage functions and measure actual net benefit. It fits the case where a busy accelerated server spends significant host CPU cycles processing storage and network control tasks. It cannot prove success here; instead the required evidence is an injected supported offload fault is localized to the DPU service or interface in cluster telemetry.
D:This observation would validate a different control: Define supported DPU management, update and rollback procedures separate from host OS changes. It fits the case where an administrator configures BlueField networking without documenting ownership of DPU software and firmware updates. It cannot prove success here; instead the required evidence is an injected supported offload fault is localized to the DPU service or interface in cluster telemetry.
Question 267
An architecture review proposes: Compare usable memory capacity, bandwidth and software support alongside compute performance. The underlying requirement is: A team chooses an accelerator based only on peak compute rating although the model nearly fills GPU memory. What technical rationale MOST strongly supports it?
- Useful GPU throughput requires enough parallel work to keep execution resources occupied.
- The fastest arithmetic unit is insufficient when the model and runtime state cannot fit in memory.
- GPU acceleration must account for transfer and launch overhead, not merely theoretical kernel throughput.
- CPUs are strong general-purpose control processors while GPUs excel at many parallel numerical operations.
Explanation
Correct Answer
B:The governing principle is that the fastest arithmetic unit is insufficient when the model and runtime state cannot fit in memory. It supports the measure to compare usable memory capacity, bandwidth and software support alongside compute performance. because a team chooses an accelerator based only on peak compute rating although the model nearly fills GPU memory. Without the control, the model can fail allocation or thrash despite the accelerator's high advertised FLOPS. The most useful acceptance evidence is: A load run records peak allocated memory, bandwidth utilization and sustained task throughput.
Incorrect Answers
A:That principle is relevant to this distinct case: A data scientist reports low GPU utilization even though thousands of tiny kernels launch every second. The associated response would be to profile kernel launch overhead, batching and occupancy before purchasing additional GPUs. It does not describe the central reason for this case, which is that the fastest arithmetic unit is insufficient when the model and runtime state cannot fit in memory.
C:That principle is relevant to this distinct case: A short GPU kernel performs little arithmetic but repeatedly copies tiny host buffers across PCIe. The associated response would be to batch transfers and minimize host-device movement where data dependence and memory budgets permit. It does not describe the central reason for this case, which is that the fastest arithmetic unit is insufficient when the model and runtime state cannot fit in memory.
D:That principle is relevant to this distinct case: A data-center technician assumes one GPU is a universal replacement for a CPU in every application. The associated response would be to keep control-flow-heavy sequential orchestration on CPUs and map suitable parallel kernels to GPUs. It does not describe the central reason for this case, which is that the fastest arithmetic unit is insufficient when the model and runtime state cannot fit in memory.
Question 268
A team omits this measure: Enforce hypervisor, guest-driver and tenant resource permissions for virtualized accelerated workloads. In this environment: An organization consolidates multiple AI tenants but has no process to constrain access to virtual GPU devices and images. Which resulting problem is MOST likely?
- Oversubscription can prevent VM startup or lead to unpredictable application performance.
- A misconfigured guest may gain access to shared device resources or another tenant's data.
- An unsupported sharing model may fail user isolation, driver compatibility or license policy.
- A tenant may receive weaker isolation or unpredictable latency than its service contract requires.
Explanation
Correct Answer
B:The primary exposure is that a misconfigured guest may gain access to shared device resources or another tenant's data. This follows from an organization consolidates multiple AI tenants but has no process to constrain access to virtual GPU devices and images. where the omitted measure was to enforce hypervisor, guest-driver and tenant resource permissions for virtualized accelerated workloads. The technical reason is that gPU virtualization requires isolation at management, device allocation and data access layers. A good operational validation after repair would be: An unauthorized guest test cannot view unassigned device resources or protected host management endpoints.
Incorrect Answers
A:This exposure arises in a different situation: A virtualization cluster admits more virtual-GPU memory assignments than can be backed by the available hardware profiles. There, the corresponding safeguard is to plan explicit virtual-GPU profile capacity and admission controls instead of assuming limitless overcommit. In the present case, the more direct risk is that a misconfigured guest may gain access to shared device resources or another tenant's data.
C:This exposure arises in a different situation: A virtual-desktop deployment requires accelerated graphics for several isolated users on shared server GPUs. There, the corresponding safeguard is to evaluate supported NVIDIA vGPU capabilities, profiles and licensing against the application requirements. In the present case, the more direct risk is that a misconfigured guest may gain access to shared device resources or another tenant's data.
D:This exposure arises in a different situation: A team proposes GPU time slicing as equivalent to hardware partitioning with independent memory fault isolation. There, the corresponding safeguard is to distinguish time-shared GPU execution from supported MIG hardware isolation before tenant placement. In the present case, the more direct risk is that a misconfigured guest may gain access to shared device resources or another tenant's data.
Question 269
The selected AI infrastructure safeguard is not implemented: Model total training-state memory and evaluate supported sharding or memory-saving strategies. The observed situation is: A team estimates LLM training memory from the file size of weights and omits optimizer and gradient storage. What exposure remains?
- The job can fail with out-of-memory errors even before the first optimization epoch completes.
- A procurement plan may understate memory by several multiples and stall the training job.
- Unfavorable placement may add host interconnect traffic and degrade distributed workload throughput.
- Distributed training may spend most of its time moving data instead of computing gradients.
Explanation
Correct Answer
B:The primary exposure is that a procurement plan may understate memory by several multiples and stall the training job. This follows from a team estimates LLM training memory from the file size of weights and omits optimizer and gradient storage. where the omitted measure was to model total training-state memory and evaluate supported sharding or memory-saving strategies. The technical reason is that optimizer and gradient states can consume substantial memory beyond inference-only model weights. A good operational validation after repair would be: An instrumented training step confirms peak optimizer, gradient and activation allocations fit capacity.
Incorrect Answers
A:This exposure arises in a different situation: A training model needs 52 GB of weights and activations but the proposed accelerator has only 24 GB of usable memory. There, the corresponding safeguard is to size accelerator memory for model parameters, activations, gradients and runtime headroom before procurement. In the present case, the more direct risk is that a procurement plan may understate memory by several multiples and stall the training job.
C:This exposure arises in a different situation: Two GPUs share a host but one is connected through a constrained PCIe path to the preferred network adapter. There, the corresponding safeguard is to inspect CPU NUMA, PCIe and GPU-to-NIC affinity before assigning communication-heavy training jobs. In the present case, the more direct risk is that a procurement plan may understate memory by several multiples and stall the training job.
D:This exposure arises in a different situation: A deep-learning training server contains several GPUs but inter-GPU copies must cross a slower host path. There, the corresponding safeguard is to select an appropriate high-bandwidth GPU interconnect topology for the planned multi-GPU communication. In the present case, the more direct risk is that a procurement plan may understate memory by several multiples and stall the training job.
Question 270
A technology review accepts this control: Separate node provisioning and lifecycle management from workload scheduling with appropriate integrations. Which observed result provides the STRONGEST evidence of effectiveness?
- A load rehearsal keeps administrative access responsive while training consumes assigned worker GPUs.
- A freshly enrolled GPU node receives the approved image and can run a small scheduled health job.
- A sampled GPU serial and NIC link map correctly to the physical rack and current firmware inventory.
- A congestion injection on the training fabric does not block authorized management access and logging.
Explanation
Correct Answer
B:The strongest proof is: A freshly enrolled GPU node receives the approved image and can run a small scheduled health job. It directly tests the safeguard to separate node provisioning and lifecycle management from workload scheduling with appropriate integrations. The control addresses this specific context: Cluster operators install a scheduler but have no standard process to provision and configure new GPU nodes. It matters because cluster management establishes consistent node state while schedulers choose where jobs execute. Without it, new nodes may remain untrusted or incompatible even if they appear available for scheduling.
Incorrect Answers
A:This observation would validate a different control: Separate management or login responsibilities from dedicated GPU compute workers. It fits the case where researchers run heavy model training directly on the only cluster login node used for submissions and admin work. It cannot prove success here; instead the required evidence is a freshly enrolled GPU node receives the approved image and can run a small scheduled health job.
C:This observation would validate a different control: Maintain an authoritative inventory linking device identifiers to node, rack, NIC, firmware and support ownership. It fits the case where engineers cannot determine which network port and GPU belongs to a rack because accelerator inventory uses manual notes. It cannot prove success here; instead the required evidence is a freshly enrolled GPU node receives the approved image and can run a small scheduled health job.
D:This observation would validate a different control: Design logical or physical management, storage and workload network separation appropriate to risk. It fits the case where a research environment shares the same traffic interface for administrative SSH, storage recovery and large all-reduce jobs. It cannot prove success here; instead the required evidence is a freshly enrolled GPU node receives the approved image and can run a small scheduled health job.