NCA - AI Infrastructure and Operations NCA-AIIO Practice Test 11
Question 100
An AI infrastructure team faces this requirement: A security architect wants infrastructure service isolation from host applications that may be compromised. Which action BEST addresses the operational constraint?
- Evaluate BlueField DPU-supported trusted infrastructure processing and its administrative trust boundaries.
- Apply least-privilege administration and protected policy management for DPU-hosted infrastructure controls.
- Assess supported DPU offload for suitable networking and storage functions and measure actual net benefit.
- Assign appropriate network, security or storage infrastructure tasks to the DPU rather than model tensor computation.
Explanation
Correct Answer
A:The appropriate response is to evaluate BlueField DPU-supported trusted infrastructure processing and its administrative trust boundaries. This is specific to the case where a security architect wants infrastructure service isolation from host applications that may be compromised. The mechanism is that dedicated infrastructure processing can separate selected network and security duties from general host code. If ignored, a compromised host may gain inappropriate control of shared infrastructure functions if boundaries are absent. Validate the choice using this targeted evidence: An access test demonstrates the host cannot alter protected DPU management settings without authorization.
Incorrect Answers
B:This alternative addresses a different requirement: A data-center design uses a DPU for security enforcement but permits every workload administrator to reconfigure its policies. It would be appropriate there to apply least-privilege administration and protected policy management for DPU-hosted infrastructure controls. But the present scenario is about a security architect wants infrastructure service isolation from host applications that may be compromised. For it, the needed control is to evaluate BlueField DPU-supported trusted infrastructure processing and its administrative trust boundaries.
C:This alternative addresses a different requirement: A busy accelerated server spends significant host CPU cycles processing storage and network control tasks. It would be appropriate there to assess supported DPU offload for suitable networking and storage functions and measure actual net benefit. But the present scenario is about a security architect wants infrastructure service isolation from host applications that may be compromised. For it, the needed control is to evaluate BlueField DPU-supported trusted infrastructure processing and its administrative trust boundaries.
D:This alternative addresses a different requirement: A platform team installs a DPU and assumes it will automatically accelerate every CUDA compute kernel. It would be appropriate there to assign appropriate network, security or storage infrastructure tasks to the DPU rather than model tensor computation. But the present scenario is about a security architect wants infrastructure service isolation from host applications that may be compromised. For it, the needed control is to evaluate BlueField DPU-supported trusted infrastructure processing and its administrative trust boundaries.
Question 101
An operations team implements: Evaluate dynamic batching within the permitted queueing delay and model-specific latency budget. Which test result would MOST convincingly show the approach works?
- An evaluation compares task accuracy and peak memory for an approved adapter versus a full-tune baseline.
- The sharded training run completes one correct optimization step with measured communication overhead.
- A benchmark plots throughput versus p95 latency and selects a batch policy meeting both limits.
- After a controlled interruption, resumed training starts from a checkpoint with consistent progress and metrics.
Explanation
Correct Answer
C:The strongest proof is: A benchmark plots throughput versus p95 latency and selects a batch policy meeting both limits. It directly tests the safeguard to evaluate dynamic batching within the permitted queueing delay and model-specific latency budget. The control addresses this specific context: An inference service receives many short requests but operates each GPU one request at a time with poor utilization. It matters because batching can improve throughput but adds queueing latency and must respect serving requirements. Without it, serving costs remain high while large batches can also cause unacceptable response latency.
Incorrect Answers
A:This observation would validate a different control: Compare parameter-efficient tuning against full-model training based on quality and available accelerator memory. It fits the case where a business fine-tunes a foundation model using a small private dataset and wants to reduce training resource expense. It cannot prove success here; instead the required evidence is a benchmark plots throughput versus p95 latency and selects a batch policy meeting both limits.
B:This observation would validate a different control: Evaluate model parallelism with appropriate tensor or pipeline partitioning and high-bandwidth interconnect needs. It fits the case where a foundation-model training plan exceeds the memory of a single accelerator even after supported precision reductions. It cannot prove success here; instead the required evidence is a benchmark plots throughput versus p95 latency and selects a batch policy meeting both limits.
D:This observation would validate a different control: Write versioned training checkpoints to resilient storage and test restart from a known consistent state. It fits the case where a multi-hour distributed model training run loses all progress when one compute node is recycled for maintenance. It cannot prove success here; instead the required evidence is a benchmark plots throughput versus p95 latency and selects a batch policy meeting both limits.
Question 102
An operations team implements: Evaluate supported Tensor Core execution and numerical precision settings with quality tests. Which test result would MOST convincingly show the approach works?
- An optimized batch or kernel fusion test improves active GPU occupancy and end-to-end throughput.
- Profiling confirms eligible Tensor Core kernels run while accuracy remains within agreed tolerance.
- A load run records peak allocated memory, bandwidth utilization and sustained task throughput.
- A profiler identifies bandwidth pressure and validates an improvement from more efficient data access.
Explanation
Correct Answer
B:The strongest proof is: Profiling confirms eligible Tensor Core kernels run while accuracy remains within agreed tolerance. It directly tests the safeguard to evaluate supported Tensor Core execution and numerical precision settings with quality tests. The control addresses this specific context: A transformer inference plan overlooks special hardware paths for accelerated matrix operations. It matters because tensor Cores accelerate suitable tensor and matrix operations when libraries and data types support them. Without it, the service may miss throughput goals by running unsupported kernels or inappropriate precision modes.
Incorrect Answers
A:This observation would validate a different control: Profile kernel launch overhead, batching and occupancy before purchasing additional GPUs. It fits the case where a data scientist reports low GPU utilization even though thousands of tiny kernels launch every second. It cannot prove success here; instead the required evidence is profiling confirms eligible Tensor Core kernels run while accuracy remains within agreed tolerance.
C:This observation would validate a different control: Compare usable memory capacity, bandwidth and software support alongside compute performance. It fits the case where a team chooses an accelerator based only on peak compute rating although the model nearly fills GPU memory. It cannot prove success here; instead the required evidence is profiling confirms eligible Tensor Core kernels run while accuracy remains within agreed tolerance.
D:This observation would validate a different control: Profile effective GPU memory bandwidth and data reuse rather than only counting theoretical FLOPS. It fits the case where a model performs memory-intensive operations with simple arithmetic and sees low compute utilization. It cannot prove success here; instead the required evidence is profiling confirms eligible Tensor Core kernels run while accuracy remains within agreed tolerance.
Question 103
A technology review accepts this control: Plan client ingress, load balancing and north-south bandwidth separately from training interconnects. Which observed result provides the STRONGEST evidence of effectiveness?
- End-to-end request tests include external network latency and backend GPU service timing.
- An all-reduce test measures steady collective latency and bandwidth under representative scale.
- A multi-rack benchmark shows job throughput and congestion stay within explicit scale targets.
- A combined stress test demonstrates acceptable collective latency during realistic checkpoint traffic.
Explanation
Correct Answer
A:The strongest proof is: End-to-end request tests include external network latency and backend GPU service timing. It directly tests the safeguard to plan client ingress, load balancing and north-south bandwidth separately from training interconnects. The control addresses this specific context: A model-serving tier delivers predictions to many remote clients but engineers optimize only GPU-to-GPU links. It matters because inference network demand includes client service response paths as well as internal accelerator traffic. Without it, a fast GPU can still fail its end-user SLO when ingress or egress is the bottleneck.
Incorrect Answers
B:This observation would validate a different control: Size east-west network bandwidth and collective latency for the intended distributed training topology. It fits the case where a multi-node transformer training job exchanges large gradients across all GPU workers after each step. It cannot prove success here; instead the required evidence is end-to-end request tests include external network latency and backend GPU service timing.
C:This observation would validate a different control: Calculate fabric bisection bandwidth and oversubscription for the planned scale-out job mix. It fits the case where a leaf-spine AI network uses heavily oversubscribed uplinks and passes only simple single-node benchmarks. It cannot prove success here; instead the required evidence is end-to-end request tests include external network latency and backend GPU service timing.
D: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 end-to-end request tests include external network latency and backend GPU service timing.
Question 104
In this situation, an AI platform runs model-serving jobs and distributed training under one global namespace with no capacity boundaries. The team does not separate workload classes with appropriate quotas, priorities and isolation rules in the scheduler. Which consequence BEST characterizes the gap?
- A scheduling success may still lead to failed CUDA initialization on a mismatched node.
- GPU pods may remain pending or start without usable devices even when hardware exists.
- A large training queue can consume resources needed for customer inference requests.
- A planned repair can unexpectedly cancel recently admitted training jobs.
Explanation
Correct Answer
C:The primary exposure is that a large training queue can consume resources needed for customer inference requests. This follows from an AI platform runs model-serving jobs and distributed training under one global namespace with no capacity boundaries. where the omitted measure was to separate workload classes with appropriate quotas, priorities and isolation rules in the scheduler. The technical reason is that workload admission policies should protect latency-sensitive serving from batch training contention. A good operational validation after repair would be: A stress test shows protected inference replicas retain assigned capacity during a training surge.
Incorrect Answers
A:This exposure arises in a different situation: GPU container images work on one host but fail on new workers because drivers and runtime integration differ. There, the corresponding safeguard is to standardize supported host driver, container toolkit and GPU Operator lifecycle on eligible worker pools. In the present case, the more direct risk is that a large training queue can consume resources needed for customer inference requests.
B:This exposure arises in a different situation: A Kubernetes AI cluster contains NVIDIA GPUs but pod resource requests are invisible to the control plane. There, the corresponding safeguard is to install and validate supported GPU drivers, NVIDIA device integration and resource advertisement. In the present case, the more direct risk is that a large training queue can consume resources needed for customer inference requests.
D:This exposure arises in a different situation: A GPU node is marked for maintenance while the scheduler continues assigning new long-running training jobs to it. There, the corresponding safeguard is to cordon or drain the node through approved scheduler workflows and coordinate checkpointable tasks. In the present case, the more direct risk is that a large training queue can consume resources needed for customer inference requests.
Question 105
An architecture review proposes: Confirm the supported GPU driver, CUDA runtime and application compatibility matrix before releasing a validated node image. The underlying requirement is: A CUDA application compiled for an older environment fails to initialize after the data center deploys a newer GPU and driver image. What technical rationale MOST strongly supports it?
- The NGC catalog helps distribute optimized AI software artifacts with documented versions and dependencies.
- Triton provides a managed inference serving layer with model backends, batching and standardized endpoints.
- Containerized AI requires both host accelerator support and an authorized device exposure path to workloads.
- GPU software compatibility requires aligning supported driver interfaces, runtime libraries and application binaries.
Explanation
Correct Answer
D:The governing principle is that gPU software compatibility requires aligning supported driver interfaces, runtime libraries and application binaries. It supports the measure to confirm the supported GPU driver, CUDA runtime and application compatibility matrix before releasing a validated node image. because a CUDA application compiled for an older environment fails to initialize after the data center deploys a newer GPU and driver image. Without the control, a production inference job may fail on startup even though the accelerator hardware passes basic health checks. The most useful acceptance evidence is: A smoke-test container successfully detects the approved GPU and executes a known CUDA kernel on the target image.
Incorrect Answers
A:That principle is relevant to this distinct case: Data scientists repeatedly download unverified deep-learning container images from personal registries before sharing GPU nodes. The associated response would be to use vetted NGC catalog images and pin approved container versions with tested driver compatibility. It does not describe the central reason for this case, which is that gPU software compatibility requires aligning supported driver interfaces, runtime libraries and application binaries.
B:That principle is relevant to this distinct case: An organization manually writes a different HTTP serving process for each trained model, making batching and health reporting inconsistent. The associated response would be to evaluate NVIDIA Triton Inference Server with suitable model repositories, batching and protocol endpoints. It does not describe the central reason for this case, which is that gPU software compatibility requires aligning supported driver interfaces, runtime libraries and application binaries.
C:That principle is relevant to this distinct case: A cluster project treats a container runtime as sufficient to expose GPUs without installing drivers or configuring GPU access. The associated response would be to install compatible NVIDIA drivers and container runtime integration, then validate device visibility inside containers. It does not describe the central reason for this case, which is that gPU software compatibility requires aligning supported driver interfaces, runtime libraries and application binaries.
Question 106
The design team faces this case: Research experiments change learning rates and seeds without recording configuration or resulting checkpoints. It adopts: Track hyperparameters, dataset versions, random seeds and artifacts in an experiment registry. What evidence BEST validates the intended outcome?
- Production metadata resolves to an approved immutable model version and archived evaluation record.
- A controlled distribution shift creates a monitored quality alert and a documented evaluation action.
- The same raw sample produces matching feature vectors in both the training and inference pipelines.
- A second approved run reproduces the expected metrics within declared nondeterministic tolerances.
Explanation
Correct Answer
D:The strongest proof is: A second approved run reproduces the expected metrics within declared nondeterministic tolerances. It directly tests the safeguard to track hyperparameters, dataset versions, random seeds and artifacts in an experiment registry. The control addresses this specific context: Research experiments change learning rates and seeds without recording configuration or resulting checkpoints. It matters because reproducible AI experiments require the parameters and input versions that produced a checkpoint. Without it, an apparent accuracy improvement cannot be attributed or recreated for production approval.
Incorrect Answers
A:This observation would validate a different control: Register immutable model versions with provenance, evaluation results and approved release transitions. It fits the case where operations cannot determine which trained model is running in each deployment because artifacts are overwritten under one label. It cannot prove success here; instead the required evidence is a second approved run reproduces the expected metrics within declared nondeterministic tolerances.
B:This observation would validate a different control: Monitor production data distribution, predictive performance and drift with assigned retraining triggers. It fits the case where a demand-forecasting model loses accuracy months after deployment as user behavior changes. It cannot prove success here; instead the required evidence is a second approved run reproduces the expected metrics within declared nondeterministic tolerances.
C:This observation would validate a different control: Version preprocessing and feature transformations with the model and verify training-serving parity. It fits the case where a feature-extraction job produces different input scales in training and the live inference service. It cannot prove success here; instead the required evidence is a second approved run reproduces the expected metrics within declared nondeterministic tolerances.
Question 107
The selected AI infrastructure safeguard is not implemented: Design appropriate loss-control, priority-flow and congestion mechanisms for the approved RoCE fabric. The observed situation is: A team deploys RoCE traffic over standard congested Ethernet with no plan to handle packet loss or pauses. What exposure remains?
- Overly aggressive pauses can propagate backpressure and disrupt otherwise healthy network flows.
- Mismatched frames can be dropped or fragmented and create hard-to-diagnose application failures.
- An inappropriate default design can yield unpredictable low-latency workload behavior.
- Packet drops and uncontrolled pause behavior can cause severe tail latency in collective operations.
Explanation
Correct Answer
D:The primary exposure is that packet drops and uncontrolled pause behavior can cause severe tail latency in collective operations. This follows from a team deploys RoCE traffic over standard congested Ethernet with no plan to handle packet loss or pauses. where the omitted measure was to design appropriate loss-control, priority-flow and congestion mechanisms for the approved RoCE fabric. The technical reason is that rDMA over Ethernet requires deliberate handling of congestion and loss for predictable performance. A good operational validation after repair would be: A controlled load test confirms correct ECN or PFC behavior and stable RDMA throughput.
Incorrect Answers
A:This exposure arises in a different situation: A storage and training fabric experiences repeated priority flow-control storms that slow unrelated traffic classes. There, the corresponding safeguard is to inspect PFC pause counters and congestion telemetry before modifying quality-of-service configuration. In the present case, the more direct risk is that packet drops and uncontrolled pause behavior can cause severe tail latency in collective operations.
B:This exposure arises in a different situation: A team sets jumbo-frame MTU on GPU servers but overlooks one switch hop in a high-speed data path. There, the corresponding safeguard is to validate path MTU consistency across endpoints, switches and relevant routed interfaces. In the present case, the more direct risk is that packet drops and uncontrolled pause behavior can cause severe tail latency in collective operations.
C:This exposure arises in a different situation: An engineer chooses between InfiniBand and Ethernet but assumes both have identical congestion and transport behavior. There, the corresponding safeguard is to document protocol-specific transport, management and performance requirements before selecting the AI fabric. In the present case, the more direct risk is that packet drops and uncontrolled pause behavior can cause severe tail latency in collective operations.
Question 108
A GPU operations plan calls for: Select supported high-speed NIC or HCA capacity aligned with expected accelerator and storage traffic. The reason is that an accelerated node uses a slow network interface although its GPU data path needs high-bandwidth peer transfers. Which principle is MOST relevant?
- Training network requirements are often driven by accelerator-to-accelerator synchronization traffic.
- Inference network demand includes client service response paths as well as internal accelerator traffic.
- NIC selection can limit effective GPU cluster throughput despite ample internal device bandwidth.
- Direct accelerator data paths can reduce host-memory copies when the complete platform supports them.
Explanation
Correct Answer
C:The governing principle is that nIC selection can limit effective GPU cluster throughput despite ample internal device bandwidth. It supports the measure to select supported high-speed NIC or HCA capacity aligned with expected accelerator and storage traffic. because an accelerated node uses a slow network interface although its GPU data path needs high-bandwidth peer transfers. Without the control, network links may saturate before the GPU completes its available compute work. The most useful acceptance evidence is: A data transfer benchmark shows per-node port throughput within the planned utilization envelope.
Incorrect Answers
A:That principle is relevant to this distinct case: A multi-node transformer training job exchanges large gradients across all GPU workers after each step. The associated response would be to size east-west network bandwidth and collective latency for the intended distributed training topology. It does not describe the central reason for this case, which is that nIC selection can limit effective GPU cluster throughput despite ample internal device bandwidth.
B:That principle is relevant to this distinct case: A model-serving tier delivers predictions to many remote clients but engineers optimize only GPU-to-GPU links. The associated response would be to plan client ingress, load balancing and north-south bandwidth separately from training interconnects. It does not describe the central reason for this case, which is that nIC selection can limit effective GPU cluster throughput despite ample internal device bandwidth.
D:That principle is relevant to this distinct case: A research team expects GPUs to access distributed data without unnecessary staging through host CPU memory. The associated response would be to evaluate supported GPUDirect RDMA and the required GPU, NIC, driver and topology compatibility. It does not describe the central reason for this case, which is that nIC selection can limit effective GPU cluster throughput despite ample internal device bandwidth.
Question 109
An architecture review proposes: Size east-west network bandwidth and collective latency for the intended distributed training topology. The underlying requirement is: A multi-node transformer training job exchanges large gradients across all GPU workers after each step. What technical rationale MOST strongly supports it?
- Training network requirements are often driven by accelerator-to-accelerator synchronization traffic.
- NIC selection can limit effective GPU cluster throughput despite ample internal device bandwidth.
- Inference network demand includes client service response paths as well as internal accelerator traffic.
- Network isolation limits unauthorized access and protects operational control from workload contention.
Explanation
Correct Answer
A:The governing principle is that training network requirements are often driven by accelerator-to-accelerator synchronization traffic. It supports the measure to size east-west network bandwidth and collective latency for the intended distributed training topology. because a multi-node transformer training job exchanges large gradients across all GPU workers after each step. Without the control, link congestion may erase expected multi-node speedups while GPU computation remains efficient. The most useful acceptance evidence is: An all-reduce test measures steady collective latency and bandwidth under representative scale.
Incorrect Answers
B:That principle is relevant to this distinct case: An accelerated node uses a slow network interface although its GPU data path needs high-bandwidth peer transfers. The associated response would be to select supported high-speed NIC or HCA capacity aligned with expected accelerator and storage traffic. It does not describe the central reason for this case, which is that training network requirements are often driven by accelerator-to-accelerator synchronization traffic.
C:That principle is relevant to this distinct case: A model-serving tier delivers predictions to many remote clients but engineers optimize only GPU-to-GPU links. The associated response would be to plan client ingress, load balancing and north-south bandwidth separately from training interconnects. It does not describe the central reason for this case, which is that training network requirements are often driven by accelerator-to-accelerator synchronization traffic.
D:That principle is relevant to this distinct case: A shared GPU cluster combines tenant inference traffic and privileged management access on one unmanaged network. The associated response would be to segment workload and management paths with required access control and monitored boundaries. It does not describe the central reason for this case, which is that training network requirements are often driven by accelerator-to-accelerator synchronization traffic.
Question 110
An operations team implements: Inspect supported NVLink status and error counters alongside application collective performance. Which test result would MOST convincingly show the approach works?
- A central dashboard shows per-GPU health, utilization and temperature with validated alert ownership.
- A high-context load test reports peak memory and triggers a controlled alert before allocation failure.
- A link-health check identifies the affected path and an approved repair restores collective time.
- A workload test shows the clock-reduction reason and validates speed recovery after thermal correction.
Explanation
Correct Answer
C:The strongest proof is: A link-health check identifies the affected path and an approved repair restores collective time. It directly tests the safeguard to inspect supported NVLink status and error counters alongside application collective performance. The control addresses this specific context: Multi-GPU jobs slow after cabling work and a topology view shows one NVLink path with elevated errors. It matters because gPU interconnect health directly influences distributed work within supported accelerator systems. Without it, a degraded link can add latency while individual GPU compute metrics look healthy.
Incorrect Answers
A:This observation would validate a different control: Use NVIDIA DCGM monitoring and supported telemetry exporters for centralized device diagnostics. It fits the case where a fleet has thousands of accelerators but operators manually run nvidia-smi on each node to identify unhealthy GPUs. It cannot prove success here; instead the required evidence is a link-health check identifies the affected path and an approved repair restores collective time.
B:This observation would validate a different control: Monitor used and available GPU memory across realistic serving contexts and concurrent workload sizes. It fits the case where a model endpoint experiences out-of-memory failures only during long prompts and concurrent sessions. It cannot prove success here; instead the required evidence is a link-health check identifies the affected path and an approved repair restores collective time.
D:This observation would validate a different control: Correlate temperature, clock state and thermal-throttling reasons during sustained workload measurements. It fits the case where an accelerator runs at reduced clock speed during a full training epoch but device utilization appears high. It cannot prove success here; instead the required evidence is a link-health check identifies the affected path and an approved repair restores collective time.