NCA - AI Infrastructure and Operations NCA-AIIO Practice Test 14
Question 130
Consider this NVIDIA AI infrastructure scenario: A multi-GPU training model requires full-GPU memory but a manager proposes partitioning every GPU into tiny MIG slices. Which design choice should be made FIRST?
- Distinguish node-local NVLink or NVSwitch scaling from networked multi-node scale-out design.
- Reserve suitable full GPU devices for workloads that need whole-device capacity or communication topology.
- Model aggregate east-west traffic and network oversubscription at the intended cluster scale.
- Use an appropriate workload scheduler with GPU resource accounting and topology-aware placement.
Explanation
Correct Answer
B:The appropriate response is to reserve suitable full GPU devices for workloads that need whole-device capacity or communication topology. This is specific to the case where a multi-GPU training model requires full-GPU memory but a manager proposes partitioning every GPU into tiny MIG slices. The mechanism is that gPU partitioning is a workload-isolation choice, not an automatic substitute for full-device training capacity. If ignored, an otherwise valid model may be unable to fit in a slice or access the intended multi-GPU topology. Validate the choice using this targeted evidence: Scheduling tests place full-device training jobs and separate eligible small inference slices correctly.
Incorrect Answers
A:This alternative addresses a different requirement: A team adds more GPUs inside one system and assumes the same link topology extends to GPUs in other racks. It would be appropriate there to distinguish node-local NVLink or NVSwitch scaling from networked multi-node scale-out design. But the present scenario is about a multi-GPU training model requires full-GPU memory but a manager proposes partitioning every GPU into tiny MIG slices. For it, the needed control is to reserve suitable full GPU devices for workloads that need whole-device capacity or communication topology.
C:This alternative addresses a different requirement: A data center scales from four to sixty-four training nodes without upgrading per-node network ports. It would be appropriate there to model aggregate east-west traffic and network oversubscription at the intended cluster scale. But the present scenario is about a multi-GPU training model requires full-GPU memory but a manager proposes partitioning every GPU into tiny MIG slices. For it, the needed control is to reserve suitable full GPU devices for workloads that need whole-device capacity or communication topology.
D:This alternative addresses a different requirement: A shared accelerator fleet has idle GPUs in some nodes while large training jobs wait without proper scheduling. It would be appropriate there to use an appropriate workload scheduler with GPU resource accounting and topology-aware placement. But the present scenario is about a multi-GPU training model requires full-GPU memory but a manager proposes partitioning every GPU into tiny MIG slices. For it, the needed control is to reserve suitable full GPU devices for workloads that need whole-device capacity or communication topology.
Question 131
The team recommends this safeguard: Calculate worst-case electrical demand and verify PDU, branch-circuit and feed ratings with facilities staff. It applies because a proposed rack contains high-density GPU servers but its power distribution units cannot deliver full simultaneous load. Which concept justifies the design?
- Local thermal conditions determine equipment behavior more directly than a remote room thermostat.
- Power budgets must include realistic concurrent server demand and approved electrical safety margins.
- Accelerator compute density makes cooling capacity a facility-level prerequisite, not a software setting.
- Thermal or power constraints can reduce effective GPU clocks without changing model code.
Explanation
Correct Answer
B:The governing principle is that power budgets must include realistic concurrent server demand and approved electrical safety margins. It supports the measure to calculate worst-case electrical demand and verify PDU, branch-circuit and feed ratings with facilities staff. because a proposed rack contains high-density GPU servers but its power distribution units cannot deliver full simultaneous load. Without the control, a circuit overload can trip a rack feed and interrupt many training jobs at once. The most useful acceptance evidence is: A facilities-approved electrical load plan meets breaker, feed and redundant-path capacity requirements.
Incorrect Answers
A: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 power budgets must include realistic concurrent server demand and approved electrical safety margins.
C:That principle is relevant to this distinct case: An existing server room can cool ordinary web racks but a planned accelerated rack doubles local heat density. The associated response would be to validate rack heat rejection and cooling solution against documented maximum system power. It does not describe the central reason for this case, which is that power budgets must include realistic concurrent server demand and approved electrical safety margins.
D:That principle is relevant to this distinct case: An operations team sees lower training throughput on warm afternoons and suspects application regressions. The associated response would be to correlate accelerator temperature, power limit and clock throttling telemetry with ambient conditions. It does not describe the central reason for this case, which is that power budgets must include realistic concurrent server demand and approved electrical safety margins.
Question 132
In this situation, an operations team has telemetry for thousands of devices but no reliable failure labels and wants to group similar behavior. The team does not explore unsupervised clustering using selected normalized features and verify cluster interpretability. Which consequence BEST characterizes the gap?
- The project may use unsuitable classification metrics and overlook unsafe exploration behavior.
- Capacity planning may ignore the different training memory and accelerator needs of a deep network.
- The team may invent labels or treat arbitrary clusters as proven hardware-failure predictions.
- The project may promise improvements from additional data even though its rule engine never learns.
Explanation
Correct Answer
C:The primary exposure is that the team may invent labels or treat arbitrary clusters as proven hardware-failure predictions. This follows from an operations team has telemetry for thousands of devices but no reliable failure labels and wants to group similar behavior. where the omitted measure was to explore unsupervised clustering using selected normalized features and verify cluster interpretability. The technical reason is that unsupervised learning discovers structure without relying on a target label for every example. A good operational validation after repair would be: A sample review establishes meaningful cluster properties without claiming the clusters predict confirmed incidents.
Incorrect Answers
A: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 the team may invent labels or treat arbitrary clusters as proven hardware-failure predictions.
B:This exposure arises in a different situation: An enterprise compares a decision tree with a deep multilayer vision network and claims both use identical learning architectures. There, the corresponding safeguard is to classify deep learning as a subset of machine learning using layered neural representations. In the present case, the more direct risk is that the team may invent labels or treat arbitrary clusters as proven hardware-failure predictions.
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 the team may invent labels or treat arbitrary clusters as proven hardware-failure predictions.
Question 133
A data-center engineer must resolve the following situation: A cluster uses redundant electrical feeds but both power cables are plugged into the same unprotected PDU. What is the MOST appropriate response?
- Validate truly independent power paths and loss-of-feed operation with qualified facility engineers.
- Coordinate workload reduction and protected shutdown windows against available cooling redundancy.
- Validate rack heat rejection and cooling solution against documented maximum system power.
- Correlate accelerator temperature, power limit and clock throttling telemetry with ambient conditions.
Explanation
Correct Answer
A:The appropriate response is to validate truly independent power paths and loss-of-feed operation with qualified facility engineers. This is specific to the case where a cluster uses redundant electrical feeds but both power cables are plugged into the same unprotected PDU. The mechanism is that redundancy only mitigates a failure if the supposedly independent paths avoid the same fault domain. If ignored, one PDU failure may still switch off both power supplies and stop the cluster. Validate the choice using this targeted evidence: A controlled failover plan demonstrates supported equipment remains powered after either approved feed is lost.
Incorrect Answers
B:This alternative addresses a different requirement: Facilities schedules maintenance to an air-handling unit while operations keeps every GPU rack at full demand. It would be appropriate there to coordinate workload reduction and protected shutdown windows against available cooling redundancy. But the present scenario is about a cluster uses redundant electrical feeds but both power cables are plugged into the same unprotected PDU. For it, the needed control is to validate truly independent power paths and loss-of-feed operation with qualified facility engineers.
C:This alternative addresses a different requirement: An existing server room can cool ordinary web racks but a planned accelerated rack doubles local heat density. It would be appropriate there to validate rack heat rejection and cooling solution against documented maximum system power. But the present scenario is about a cluster uses redundant electrical feeds but both power cables are plugged into the same unprotected PDU. For it, the needed control is to validate truly independent power paths and loss-of-feed operation with qualified facility engineers.
D:This alternative addresses a different requirement: An operations team sees lower training throughput on warm afternoons and suspects application regressions. It would be appropriate there to correlate accelerator temperature, power limit and clock throttling telemetry with ambient conditions. But the present scenario is about a cluster uses redundant electrical feeds but both power cables are plugged into the same unprotected PDU. For it, the needed control is to validate truly independent power paths and loss-of-feed operation with qualified facility engineers.
Question 134
An architecture review proposes: Keep control-flow-heavy sequential orchestration on CPUs and map suitable parallel kernels to GPUs. The underlying requirement is: A data-center technician assumes one GPU is a universal replacement for a CPU in every application. What technical rationale MOST strongly supports it?
- CPUs are strong general-purpose control processors while GPUs excel at many parallel numerical operations.
- The fastest arithmetic unit is insufficient when the model and runtime state cannot fit in memory.
- Reduced precision can accelerate supported computations but may change numerical stability and accuracy.
- Useful GPU throughput requires enough parallel work to keep execution resources occupied.
Explanation
Correct Answer
A:The governing principle is that cPUs are strong general-purpose control processors while GPUs excel at many parallel numerical operations. It supports the measure to keep control-flow-heavy sequential orchestration on CPUs and map suitable parallel kernels to GPUs. because a data-center technician assumes one GPU is a universal replacement for a CPU in every application. Without the control, small serial tasks can incur overhead or run inefficiently when forced onto accelerators. The most useful acceptance evidence is: A profile shows sequential coordination remains CPU-resident while parallel compute is offloaded appropriately.
Incorrect Answers
B:That principle is relevant to this distinct case: A team chooses an accelerator based only on peak compute rating although the model nearly fills GPU memory. The associated response would be to compare usable memory capacity, bandwidth and software support alongside compute performance. It does not describe the central reason for this case, which is that cPUs are strong general-purpose control processors while GPUs excel at many parallel numerical operations.
C:That principle is relevant to this distinct case: A team reduces model precision to boost GPU throughput without assessing the resulting numerical behavior. The associated response would be to benchmark supported FP16, BF16 or other lower-precision modes against the application quality threshold. It does not describe the central reason for this case, which is that cPUs are strong general-purpose control processors while GPUs excel at many parallel numerical operations.
D: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 cPUs are strong general-purpose control processors while GPUs excel at many parallel numerical operations.
Question 135
A data-center manager declines to plan explicit virtual-GPU profile capacity and admission controls instead of assuming limitless overcommit. The operational context is: A virtualization cluster admits more virtual-GPU memory assignments than can be backed by the available hardware profiles. Which risk is MOST direct?
- An unsupported sharing model may fail user isolation, driver compatibility or license policy.
- A misconfigured guest may gain access to shared device resources or another tenant's data.
- Oversubscription can prevent VM startup or lead to unpredictable application performance.
- Workloads may lose accelerated features or support coverage even when a GPU is physically present.
Explanation
Correct Answer
C:The primary exposure is that oversubscription can prevent VM startup or lead to unpredictable application performance. This follows from a virtualization cluster admits more virtual-GPU memory assignments than can be backed by the available hardware profiles. where the omitted measure was to plan explicit virtual-GPU profile capacity and admission controls instead of assuming limitless overcommit. The technical reason is that accelerator virtualization remains constrained by physical memory and supported profile limits. A good operational validation after repair would be: A capacity test rejects excessive profile requests and admits a safe mix under documented limits.
Incorrect Answers
A: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 oversubscription can prevent VM startup or lead to unpredictable application performance.
B:This exposure arises in a different situation: An organization consolidates multiple AI tenants but has no process to constrain access to virtual GPU devices and images. There, the corresponding safeguard is to enforce hypervisor, guest-driver and tenant resource permissions for virtualized accelerated workloads. In the present case, the more direct risk is that oversubscription can prevent VM startup or lead to unpredictable application performance.
D:This exposure arises in a different situation: An IT department deploys vGPU workloads but ignores required enterprise license and guest-driver version alignment. There, the corresponding safeguard is to verify supported vGPU software version compatibility and the applicable licensing lifecycle. In the present case, the more direct risk is that oversubscription can prevent VM startup or lead to unpredictable application performance.
Question 136
A technology review accepts this control: Match NIM, Triton, and model customization components to serving requirements and supported integration boundaries. Which observed result provides the STRONGEST evidence of effectiveness?
- A pilot completes an approved training job while access, data controls and billing match requirements.
- A pilot confirms the intended offload function operates while host CPU load and management access are measured.
- A test node receives the approved image and appears consistently in the inventory and health dashboard.
- A capability matrix maps the training, serving and evaluation responsibilities to supported selected components.
Explanation
Correct Answer
D:The strongest proof is: A capability matrix maps the training, serving and evaluation responsibilities to supported selected components. It directly tests the safeguard to match NIM, Triton, and model customization components to serving requirements and supported integration boundaries. The control addresses this specific context: An AI platform team wants to use packaged enterprise inference models but assumes every NVIDIA solution is interchangeable. It matters because nVIDIA solutions address different stages of the AI lifecycle and must be chosen for the specific function. Without it, a custom training tool may be incorrectly deployed as a production serving platform without required controls.
Incorrect Answers
A:This observation would validate a different control: Evaluate NVIDIA DGX Cloud offerings and supported access, security and data-placement conditions. It fits the case where a company wants an AI-development environment delivered as a managed cloud service instead of maintaining a local GPU cluster. It cannot prove success here; instead the required evidence is a capability matrix maps the training, serving and evaluation responsibilities to supported selected components.
B:This observation would validate a different control: Assess NVIDIA BlueField DPU functions for supported networking, storage and infrastructure offload. It fits the case where an operations team wants to isolate infrastructure networking and security tasks away from host application CPUs. It cannot prove success here; instead the required evidence is a capability matrix maps the training, serving and evaluation responsibilities to supported selected components.
C:This observation would validate a different control: Consider NVIDIA Base Command Manager for supported cluster provisioning, management and monitoring workflows. It fits the case where a data-center operator manages many accelerated nodes by hand and has inconsistent provisioning state across racks. It cannot prove success here; instead the required evidence is a capability matrix maps the training, serving and evaluation responsibilities to supported selected components.
Question 137
A technology review accepts this control: Select an appropriate high-bandwidth GPU interconnect topology for the planned multi-GPU communication. Which observed result provides the STRONGEST evidence of effectiveness?
- A full input-path benchmark records steady sample feed rate and GPU wait time under concurrent training.
- A sizing report attributes measured step time to compute, communication and input-data stages.
- A topology-aware bandwidth test and training profile show acceptable GPU-to-GPU transfer time.
- The data-loader benchmark maintains intended batch supply without swap pressure or CPU queue buildup.
Explanation
Correct Answer
C:The strongest proof is: A topology-aware bandwidth test and training profile show acceptable GPU-to-GPU transfer time. It directly tests the safeguard to select an appropriate high-bandwidth GPU interconnect topology for the planned multi-GPU communication. The control addresses this specific context: A deep-learning training server contains several GPUs but inter-GPU copies must cross a slower host path. It matters because inter-GPU bandwidth and topology affect scaling when tensors are exchanged frequently. Without it, distributed training may spend most of its time moving data instead of computing gradients.
Incorrect Answers
A:This observation would validate a different control: Plan storage throughput, metadata operations and prefetch behavior using representative input data. It fits the case where a vision training job reads millions of small images from one slow network file share and leaves accelerators waiting. It cannot prove success here; instead the required evidence is a topology-aware bandwidth test and training profile show acceptable GPU-to-GPU transfer time.
B:This observation would validate a different control: Benchmark representative iteration time, memory use and input-data rate before finalizing hardware sizing. It fits the case where a platform engineer sizes a new cluster from vendor peak GPU FLOPS without profiling the actual training pipeline. It cannot prove success here; instead the required evidence is a topology-aware bandwidth test and training profile show acceptable GPU-to-GPU transfer time.
D:This observation would validate a different control: Size host RAM and CPU data-loader workers for preprocessing, buffering and model-feed requirements. It fits the case where a GPU server repeatedly runs out of host memory when constructing large mini-batches from many source files. It cannot prove success here; instead the required evidence is a topology-aware bandwidth test and training profile show acceptable GPU-to-GPU transfer time.
Question 138
The following safeguard was deployed: Select supported high-speed NIC or HCA capacity aligned with expected accelerator and storage traffic. What operational observation would demonstrate its success?
- 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.
- A combined stress test demonstrates acceptable collective latency during realistic checkpoint traffic.
- An all-reduce test measures steady collective latency and bandwidth under representative scale.
Explanation
Correct Answer
A:The strongest proof is: A data transfer benchmark shows per-node port throughput within the planned utilization envelope. It directly tests the safeguard to select supported high-speed NIC or HCA capacity aligned with expected accelerator and storage traffic. The control addresses this specific context: An accelerated node uses a slow network interface although its GPU data path needs high-bandwidth peer transfers. It matters because nIC selection can limit effective GPU cluster throughput despite ample internal device bandwidth. Without it, network links may saturate before the GPU completes its available compute work.
Incorrect Answers
B:This observation would validate a different control: Segment workload and management paths with required access control and monitored boundaries. It fits the case where a shared GPU cluster combines tenant inference traffic and privileged management access on one unmanaged network. It cannot prove success here; instead the required evidence is a data transfer benchmark shows per-node port throughput within the planned utilization envelope.
C:This observation would validate a different control: Estimate concurrent network needs and evaluate storage-traffic isolation or quality-of-service policies. It fits the case where cluster operators run training dataset reads and checkpoint writes through the same fabric as synchronous gradient traffic. It cannot prove success here; instead the required evidence is a data transfer benchmark shows per-node port throughput within the planned utilization envelope.
D: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 data transfer benchmark shows per-node port throughput within the planned utilization envelope.
Question 139
A team omits this measure: Evaluate supported GPUDirect RDMA and the required GPU, NIC, driver and topology compatibility. In this environment: A research team expects GPUs to access distributed data without unnecessary staging through host CPU memory. Which resulting problem is MOST likely?
- Checkpoint bursts may stall synchronized GPU workers and extend training job duration.
- Unsupported assumptions may cause fallback transfers and unexpected CPU overhead.
- A fast GPU can still fail its end-user SLO when ingress or egress is the bottleneck.
- Network links may saturate before the GPU completes its available compute work.
Explanation
Correct Answer
B:The primary exposure is that unsupported assumptions may cause fallback transfers and unexpected CPU overhead. This follows from a research team expects GPUs to access distributed data without unnecessary staging through host CPU memory. where the omitted measure was to evaluate supported GPUDirect RDMA and the required GPU, NIC, driver and topology compatibility. The technical reason is that direct accelerator data paths can reduce host-memory copies when the complete platform supports them. A good operational validation after repair would be: A validated supported path records lower host-copy overhead with correct end-to-end data delivery.
Incorrect Answers
A:This exposure arises in a different situation: Cluster operators run training dataset reads and checkpoint writes through the same fabric as synchronous gradient traffic. There, the corresponding safeguard is to estimate concurrent network needs and evaluate storage-traffic isolation or quality-of-service policies. In the present case, the more direct risk is that unsupported assumptions may cause fallback transfers and unexpected CPU overhead.
C:This exposure arises in a different situation: A model-serving tier delivers predictions to many remote clients but engineers optimize only GPU-to-GPU links. There, the corresponding safeguard is to plan client ingress, load balancing and north-south bandwidth separately from training interconnects. In the present case, the more direct risk is that unsupported assumptions may cause fallback transfers and unexpected CPU overhead.
D: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 unsupported assumptions may cause fallback transfers and unexpected CPU overhead.
Question 140
An organization is evaluating an AI deployment: A recommendation project improves after collecting representative interaction logs but management attributes all gains to GPU upgrades. What solution MOST directly fits the stated need?
- Identify accessible accelerated computation as an enabler of feasible large-scale model training.
- Include interconnect performance among the conditions supporting modern distributed AI training progress.
- Use transfer learning appropriately and validate downstream quality with the domain-specific dataset.
- Evaluate data quantity, label quality and dataset representativeness alongside accelerated hardware improvements.
Explanation
Correct Answer
D:The appropriate response is to evaluate data quantity, label quality and dataset representativeness alongside accelerated hardware improvements. This is specific to the case where a recommendation project improves after collecting representative interaction logs but management attributes all gains to GPU upgrades. The mechanism is that recent AI improvements depend on useful training data as well as computing resources and algorithms. If ignored, hardware purchases may deliver little benefit if the data remain biased or insufficient for the target users. Validate the choice using this targeted evidence: A controlled model comparison isolates the impact of data improvements from compute changes.
Incorrect Answers
A:This alternative addresses a different requirement: A research group trains larger neural networks after acquiring accelerators with much higher parallel throughput. It would be appropriate there to identify accessible accelerated computation as an enabler of feasible large-scale model training. But the present scenario is about a recommendation project improves after collecting representative interaction logs but management attributes all gains to GPU upgrades. For it, the needed control is to evaluate data quantity, label quality and dataset representativeness alongside accelerated hardware improvements.
B:This alternative addresses a different requirement: A multi-node project scales after faster network fabrics reduce the time workers spend synchronizing gradients. It would be appropriate there to include interconnect performance among the conditions supporting modern distributed AI training progress. But the present scenario is about a recommendation project improves after collecting representative interaction logs but management attributes all gains to GPU upgrades. For it, the needed control is to evaluate data quantity, label quality and dataset representativeness alongside accelerated hardware improvements.
C:This alternative addresses a different requirement: An applied AI team starts from an existing pretrained vision model rather than training its representation from random weights. It would be appropriate there to use transfer learning appropriately and validate downstream quality with the domain-specific dataset. But the present scenario is about a recommendation project improves after collecting representative interaction logs but management attributes all gains to GPU upgrades. For it, the needed control is to evaluate data quantity, label quality and dataset representativeness alongside accelerated hardware improvements.