NCA - AI Infrastructure and Operations NCA-AIIO Practice Test 30
Question 290
The design team faces this case: A multi-node transformer training job exchanges large gradients across all GPU workers after each step. It adopts: Size east-west network bandwidth and collective latency for the intended distributed training topology. What evidence BEST validates the intended outcome?
- 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.
- A multi-rack benchmark shows job throughput and congestion stay within explicit scale targets.
- A validated supported path records lower host-copy overhead with correct end-to-end data delivery.
Explanation
Correct Answer
B:The strongest proof is: An all-reduce test measures steady collective latency and bandwidth under representative scale. It directly tests the safeguard to size east-west network bandwidth and collective latency for the intended distributed training topology. The control addresses this specific context: A multi-node transformer training job exchanges large gradients across all GPU workers after each step. It matters because training network requirements are often driven by accelerator-to-accelerator synchronization traffic. Without it, link congestion may erase expected multi-node speedups while GPU computation remains efficient.
Incorrect Answers
A: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 an all-reduce test measures steady collective latency and bandwidth under representative scale.
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 an all-reduce test measures steady collective latency and bandwidth under representative scale.
D:This observation would validate a different control: Evaluate supported GPUDirect RDMA and the required GPU, NIC, driver and topology compatibility. It fits the case where a research team expects GPUs to access distributed data without unnecessary staging through host CPU memory. It cannot prove success here; instead the required evidence is an all-reduce test measures steady collective latency and bandwidth under representative scale.
Question 291
A GPU operations plan calls for: Select stable serving capacity and an appropriate lower-cost interruptible training tier with checkpoint recovery. The reason is that an AI service has predictable daily demand but a training task can be interrupted and resumed later. Which principle is MOST relevant?
- Workload criticality and recoverability should shape capacity purchasing choices.
- Physical ownership transfers hardware and operations responsibilities to the organization.
- Hybrid AI performance depends on data locality and network movement as well as rented GPU speed.
- Hybrid deployments can separate training and serving but still require trustworthy provenance and data governance.
Explanation
Correct Answer
A:The governing principle is that workload criticality and recoverability should shape capacity purchasing choices. It supports the measure to select stable serving capacity and an appropriate lower-cost interruptible training tier with checkpoint recovery. because an AI service has predictable daily demand but a training task can be interrupted and resumed later. Without the control, placing latency-critical inference on interruptible machines may create unacceptable service disruption. The most useful acceptance evidence is: Capacity and interruption tests show serving stays available while resumable training tolerates a reclaimed worker.
Incorrect Answers
B:That principle is relevant to this distinct case: A startup wants a fully owned GPU platform but lacks specialists to operate switching, cooling and firmware updates. The associated response would be to include operational staffing and maintenance lifecycle in the on-premises versus managed-cloud decision. It does not describe the central reason for this case, which is that workload criticality and recoverability should shape capacity purchasing choices.
C:That principle is relevant to this distinct case: A hybrid project keeps raw data on site but wants to train in the cloud using repeated full-dataset uploads. The associated response would be to estimate WAN transfer windows and evaluate approved data staging or local preprocessing to reduce movement. It does not describe the central reason for this case, which is that workload criticality and recoverability should shape capacity purchasing choices.
D:That principle is relevant to this distinct case: A regulated enterprise trains models locally but wants cloud inference for geographically distributed customers. The associated response would be to design a hybrid model lifecycle with approved artifact replication and separate sensitive-data access rules. It does not describe the central reason for this case, which is that workload criticality and recoverability should shape capacity purchasing choices.
Question 292
A GPU operations plan calls for: Evaluate model parallelism with appropriate tensor or pipeline partitioning and high-bandwidth interconnect needs. The reason is that a foundation-model training plan exceeds the memory of a single accelerator even after supported precision reductions. Which principle is MOST relevant?
- Model parallel training spreads different portions of a model or computation across cooperating accelerators.
- Autoregressive inference memory includes per-session runtime state beyond the static model parameter footprint.
- Inference scale-out must include model loading and initialization time, not only GPU allocation time.
- Inference runs a trained model for predictions and is often governed by response-time service-level objectives.
Explanation
Correct Answer
A:The governing principle is that model parallel training spreads different portions of a model or computation across cooperating accelerators. It supports the measure to evaluate model parallelism with appropriate tensor or pipeline partitioning and high-bandwidth interconnect needs. because a foundation-model training plan exceeds the memory of a single accelerator even after supported precision reductions. Without the control, a single device cannot hold all required model states and distributed steps may become communication bound. The most useful acceptance evidence is: The sharded training run completes one correct optimization step with measured communication overhead.
Incorrect Answers
B:That principle is relevant to this distinct case: A chat model serves increasingly long conversations and runs out of GPU memory despite stable model weights. The associated response would be to account for inference KV-cache growth and concurrency while selecting context limits and serving memory budgets. It does not describe the central reason for this case, which is that model parallel training spreads different portions of a model or computation across cooperating accelerators.
C:That principle is relevant to this distinct case: An autoscaled text-generation endpoint starts new model replicas only after a demand spike and sees long cold starts. The associated response would be to preload supported model artifacts or maintain warm capacity based on startup latency and expected demand. It does not describe the central reason for this case, which is that model parallel training spreads different portions of a model or computation across cooperating accelerators.
D:That principle is relevant to this distinct case: A hospital uses a deployed vision model to classify imaging requests within a fixed response time during business hours. The associated response would be to size online inference for bounded tail latency, concurrent sessions and availability at the actual model size. It does not describe the central reason for this case, which is that model parallel training spreads different portions of a model or computation across cooperating accelerators.
Question 293
In this situation, a GPU hall relies on one unprotected utility feed although the business requires continued operation through feed faults. The team does not assess electrical redundancy, UPS autonomy and generator integration against the site's availability objective. Which consequence BEST characterizes the gap?
- An overloaded floor or unsuitable raised-floor design could endanger personnel and equipment.
- A leak or humidity excursion may damage systems before a routine service inspection.
- A short upstream feed interruption can stop every accelerator node in the hall.
- Technicians may be unable to replace failed components without interrupting adjacent equipment.
Explanation
Correct Answer
C:The primary exposure is that a short upstream feed interruption can stop every accelerator node in the hall. This follows from a GPU hall relies on one unprotected utility feed although the business requires continued operation through feed faults. where the omitted measure was to assess electrical redundancy, UPS autonomy and generator integration against the site's availability objective. The technical reason is that facility availability depends on the complete power chain, not only server dual power supplies. A good operational validation after repair would be: A qualified power-chain review maps independent paths and documented recovery actions for feed loss.
Incorrect Answers
A:This exposure arises in a different situation: A facility plans high-density GPU racks on an upper floor without confirming structural load limits. There, the corresponding safeguard is to obtain qualified structural review of populated rack weight and floor-loading constraints. In the present case, the more direct risk is that a short upstream feed interruption can stop every accelerator node in the hall.
B:This exposure arises in a different situation: A new AI room has reliable cooling capacity but no monitoring for localized humidity or leak alarms. There, the corresponding safeguard is to add appropriate environmental sensors and integrate alerts with owned operational response procedures. In the present case, the more direct risk is that a short upstream feed interruption can stop every accelerator node in the hall.
D:This exposure arises in a different situation: A data center orders accelerated racks with rear-door heat exchangers but leaves insufficient service clearance. There, the corresponding safeguard is to check rack access, airflow and maintenance clearances before the final deployment layout. In the present case, the more direct risk is that a short upstream feed interruption can stop every accelerator node in the hall.
Question 294
A team omits this measure: Correlate per-GPU compute, memory, scheduling contention and request tail latency. In this environment: An inference service meets average latency but occasional requests become slow when multiple GPU tenants share the node. Which resulting problem is MOST likely?
- Treating every Xid as the same defect can lead to incorrect recovery and needless node replacement.
- A shared accelerator may meet aggregate throughput while violating individual service guarantees.
- An emerging device fault may remain undetected until a later uncorrectable error disrupts training.
- Replacing a healthy accelerator may leave an input-pipeline stall unresolved.
Explanation
Correct Answer
B:The primary exposure is that a shared accelerator may meet aggregate throughput while violating individual service guarantees. This follows from an inference service meets average latency but occasional requests become slow when multiple GPU tenants share the node. where the omitted measure was to correlate per-GPU compute, memory, scheduling contention and request tail latency. The technical reason is that averages obscure contention and intermittent device saturation affecting a subset of customers. A good operational validation after repair would be: Per-tenant load tests link p95 latency spikes to GPU resource contention and validate an isolation change.
Incorrect Answers
A:This exposure arises in a different situation: A containerized training job unexpectedly exits and system logs contain GPU Xid error messages. There, the corresponding safeguard is to correlate NVIDIA Xid events with driver status, hardware health and the affected job timeline. In the present case, the more direct risk is that a shared accelerator may meet aggregate throughput while violating individual service guarantees.
C:This exposure arises in a different situation: A GPU node has increasing corrected memory errors but operations waits for application crashes before investigating. There, the corresponding safeguard is to collect supported ECC error counters and evaluate hardware-health escalation thresholds. In the present case, the more direct risk is that a shared accelerator may meet aggregate throughput while violating individual service guarantees.
D:This exposure arises in a different situation: A GPU has low reported utilization while its training job is slow and the team assumes the accelerator is defective. There, the corresponding safeguard is to correlate SM utilization with CPU data loading, communication time and memory bandwidth measures. In the present case, the more direct risk is that a shared accelerator may meet aggregate throughput while violating individual service guarantees.
Question 295
In this situation, a GPU has low reported utilization while its training job is slow and the team assumes the accelerator is defective. The team does not correlate SM utilization with CPU data loading, communication time and memory bandwidth measures. Which consequence BEST characterizes the gap?
- Training can miss delivery targets while dashboards misinterpret throttled hardware as healthy.
- Replacing a healthy accelerator may leave an input-pipeline stall unresolved.
- The team may buy a faster GPU while the host-to-device data path remains saturated.
- The service can fail during demand peaks although idle memory looks sufficient.
Explanation
Correct Answer
B:The primary exposure is that replacing a healthy accelerator may leave an input-pipeline stall unresolved. This follows from a GPU has low reported utilization while its training job is slow and the team assumes the accelerator is defective. where the omitted measure was to correlate SM utilization with CPU data loading, communication time and memory bandwidth measures. The technical reason is that gPU utilization is one diagnostic indicator and does not uniquely identify the performance bottleneck. A good operational validation after repair would be: A profile explains low active GPU time through measured data-loader or collective wait intervals.
Incorrect Answers
A:This exposure arises in a different situation: An accelerator runs at reduced clock speed during a full training epoch but device utilization appears high. There, the corresponding safeguard is to correlate temperature, clock state and thermal-throttling reasons during sustained workload measurements. In the present case, the more direct risk is that replacing a healthy accelerator may leave an input-pipeline stall unresolved.
C:This exposure arises in a different situation: GPU utilization looks normal but a pipeline spends much of each step copying tensors through a constrained PCIe link. There, the corresponding safeguard is to measure host-device transfer duration, PCIe utilization and data-path topology during the real workload. In the present case, the more direct risk is that replacing a healthy accelerator may leave an input-pipeline stall unresolved.
D:This exposure arises in a different situation: A model endpoint experiences out-of-memory failures only during long prompts and concurrent sessions. There, the corresponding safeguard is to monitor used and available GPU memory across realistic serving contexts and concurrent workload sizes. In the present case, the more direct risk is that replacing a healthy accelerator may leave an input-pipeline stall unresolved.
Question 296
Consider this NVIDIA AI infrastructure scenario: A research environment shares the same traffic interface for administrative SSH, storage recovery and large all-reduce jobs. Which design choice should be made FIRST?
- Design logical or physical management, storage and workload network separation appropriate to risk.
- Maintain an authoritative inventory linking device identifiers to node, rack, NIC, firmware and support ownership.
- Map CPU, GPU, NIC, storage and interconnect components to the complete data path and workload profile.
- Identify and protect shared control-plane and storage failure domains with appropriate redundancy.
Explanation
Correct Answer
A:The appropriate response is to design logical or physical management, storage and workload network separation appropriate to risk. This is specific to the case where a research environment shares the same traffic interface for administrative SSH, storage recovery and large all-reduce jobs. The mechanism is that isolation of control traffic prevents data-plane congestion from blocking administrative response. If ignored, an intense training collective can make the node-management plane unavailable during an outage. Validate the choice using this targeted evidence: A congestion injection on the training fabric does not block authorized management access and logging.
Incorrect Answers
B:This alternative addresses a different requirement: Engineers cannot determine which network port and GPU belongs to a rack because accelerator inventory uses manual notes. It would be appropriate there to maintain an authoritative inventory linking device identifiers to node, rack, NIC, firmware and support ownership. But the present scenario is about a research environment shares the same traffic interface for administrative SSH, storage recovery and large all-reduce jobs. For it, the needed control is to design logical or physical management, storage and workload network separation appropriate to risk.
C:This alternative addresses a different requirement: An accelerated server specification lists GPUs and CPUs but fails to allocate suitable network adapters or high-bandwidth storage. It would be appropriate there to map CPU, GPU, NIC, storage and interconnect components to the complete data path and workload profile. But the present scenario is about a research environment shares the same traffic interface for administrative SSH, storage recovery and large all-reduce jobs. For it, the needed control is to design logical or physical management, storage and workload network separation appropriate to risk.
D:This alternative addresses a different requirement: A cluster stores all training metadata and management state on one server without recovery copies. It would be appropriate there to identify and protect shared control-plane and storage failure domains with appropriate redundancy. But the present scenario is about a research environment shares the same traffic interface for administrative SSH, storage recovery and large all-reduce jobs. For it, the needed control is to design logical or physical management, storage and workload network separation appropriate to risk.
Question 297
An AI infrastructure team faces this requirement: An inference service receives many short requests but operates each GPU one request at a time with poor utilization. Which action BEST addresses the operational constraint?
- Compare parameter-efficient tuning against full-model training based on quality and available accelerator memory.
- Evaluate model parallelism with appropriate tensor or pipeline partitioning and high-bandwidth interconnect needs.
- Optimize for sustained training throughput, memory capacity and batch processing rather than single-request latency.
- Evaluate dynamic batching within the permitted queueing delay and model-specific latency budget.
Explanation
Correct Answer
D:The appropriate response is to evaluate dynamic batching within the permitted queueing delay and model-specific latency budget. This is specific to the case where an inference service receives many short requests but operates each GPU one request at a time with poor utilization. The mechanism is that batching can improve throughput but adds queueing latency and must respect serving requirements. If ignored, serving costs remain high while large batches can also cause unacceptable response latency. Validate the choice using this targeted evidence: A benchmark plots throughput versus p95 latency and selects a batch policy meeting both limits.
Incorrect Answers
A:This alternative addresses a different requirement: A business fine-tunes a foundation model using a small private dataset and wants to reduce training resource expense. It would be appropriate there to compare parameter-efficient tuning against full-model training based on quality and available accelerator memory. But the present scenario is about an inference service receives many short requests but operates each GPU one request at a time with poor utilization. For it, the needed control is to evaluate dynamic batching within the permitted queueing delay and model-specific latency budget.
B:This alternative addresses a different requirement: A foundation-model training plan exceeds the memory of a single accelerator even after supported precision reductions. It would be appropriate there to evaluate model parallelism with appropriate tensor or pipeline partitioning and high-bandwidth interconnect needs. But the present scenario is about an inference service receives many short requests but operates each GPU one request at a time with poor utilization. For it, the needed control is to evaluate dynamic batching within the permitted queueing delay and model-specific latency budget.
C:This alternative addresses a different requirement: A research team trains a large image classifier nightly with lengthy epochs and no strict per-request response deadline. It would be appropriate there to optimize for sustained training throughput, memory capacity and batch processing rather than single-request latency. But the present scenario is about an inference service receives many short requests but operates each GPU one request at a time with poor utilization. For it, the needed control is to evaluate dynamic batching within the permitted queueing delay and model-specific latency budget.
Question 298
The selected AI infrastructure safeguard is not implemented: Associate GPU and chassis thermal events with rack location and facility environmental telemetry. The observed situation is: A training room experiences repeated thermal warnings while hardware inventory lacks rack and location mapping. What exposure remains?
- One failed redundant component can remain unnoticed until a second fault causes a full outage.
- Unplanned repairs may destroy unsaved training state or block critical inference workloads.
- Operators may replace healthy GPUs while leaving the actual localized cooling problem unresolved.
- A false alarm could drain healthy nodes and interrupt priority training across the cluster.
Explanation
Correct Answer
C:The primary exposure is that operators may replace healthy GPUs while leaving the actual localized cooling problem unresolved. This follows from a training room experiences repeated thermal warnings while hardware inventory lacks rack and location mapping. where the omitted measure was to associate GPU and chassis thermal events with rack location and facility environmental telemetry. The technical reason is that operational diagnosis improves when accelerator sensors are correlated with their physical cooling context. A good operational validation after repair would be: An alert joins a high GPU temperature with its rack inlet sensor and confirms the response owner.
Incorrect Answers
A:This exposure arises in a different situation: Data-center management sees correct fan speeds but receives no alert when an accelerator power module reports a fault. There, the corresponding safeguard is to integrate supported power-supply and chassis fault signals with prioritized alert routing. In the present case, the more direct risk is that operators may replace healthy GPUs while leaving the actual localized cooling problem unresolved.
B:This exposure arises in a different situation: Technicians take shared GPU nodes offline during business hours without coordinating active checkpointing jobs. There, the corresponding safeguard is to enforce maintenance windows with scheduler drain, workload notice and readiness verification. In the present case, the more direct risk is that operators may replace healthy GPUs while leaving the actual localized cooling problem unresolved.
D:This exposure arises in a different situation: Engineers trigger automated node remediation from broad alerts without checking device identities or active jobs. There, the corresponding safeguard is to scope remediation to verified affected nodes and require approval for destructive maintenance actions. In the present case, the more direct risk is that operators may replace healthy GPUs while leaving the actual localized cooling problem unresolved.
Question 299
The selected AI infrastructure safeguard is not implemented: Validate rack heat rejection and cooling solution against documented maximum system power. The observed situation is: An existing server room can cool ordinary web racks but a planned accelerated rack doubles local heat density. What exposure remains?
- One chassis can throttle or shut down without a building-wide temperature alarm.
- Hotspots can cause thermal throttling or emergency shutdown while overall room temperature seems normal.
- A circuit overload can trip a rack feed and interrupt many training jobs at once.
- Cooling interruptions or leaks can make new systems unusable or risk equipment damage.
Explanation
Correct Answer
B:The primary exposure is that hotspots can cause thermal throttling or emergency shutdown while overall room temperature seems normal. This follows from an existing server room can cool ordinary web racks but a planned accelerated rack doubles local heat density. where the omitted measure was to validate rack heat rejection and cooling solution against documented maximum system power. The technical reason is that accelerator compute density makes cooling capacity a facility-level prerequisite, not a software setting. A good operational validation after repair would be: Measured inlet temperatures and predicted rack heat remain inside approved vendor and facility limits.
Incorrect Answers
A:This exposure arises in a different situation: A rack runs multiple powerful accelerators but operators monitor only average room temperature, not intake air. There, the corresponding safeguard is to collect rack inlet and component thermal measurements during representative peak workloads. In the present case, the more direct risk is that hotspots can cause thermal throttling or emergency shutdown while overall room temperature seems normal.
C:This exposure arises in a different situation: A proposed rack contains high-density GPU servers but its power distribution units cannot deliver full simultaneous load. There, the corresponding safeguard is to calculate worst-case electrical demand and verify PDU, branch-circuit and feed ratings with facilities staff. In the present case, the more direct risk is that hotspots can cause thermal throttling or emergency shutdown while overall room temperature seems normal.
D:This exposure arises in a different situation: A team wants to deploy liquid-cooled GPU systems in a building that has no validated coolant distribution network. There, the corresponding safeguard is to check facility liquid cooling availability, heat exchangers, leak detection and vendor flow requirements. In the present case, the more direct risk is that hotspots can cause thermal throttling or emergency shutdown while overall room temperature seems normal.
Question 300
The team recommends this safeguard: Validate approved facility life-safety procedures and equipment integration with qualified engineers. It applies because facility operations adds high-density racks without reviewing fire detection and emergency shutoff responsibilities. Which concept justifies the design?
- Fire detection and safety systems must be planned alongside accelerator and cooling deployment.
- Physical maintainability and cooling access are mandatory facility design constraints.
- Facility monitoring should detect environmental conditions that threaten accelerated equipment operation.
- Facility cabling must respect physical limits and support reliable future fault identification.
Explanation
Correct Answer
A:The governing principle is that fire detection and safety systems must be planned alongside accelerator and cooling deployment. It supports the measure to validate approved facility life-safety procedures and equipment integration with qualified engineers. because facility operations adds high-density racks without reviewing fire detection and emergency shutoff responsibilities. Without the control, inadequate emergency procedures can expose people and equipment to preventable hazards. The most useful acceptance evidence is: The facility sign-off confirms appropriate detection, response ownership and approved emergency controls.
Incorrect Answers
B:That principle is relevant to this distinct case: A data center orders accelerated racks with rear-door heat exchangers but leaves insufficient service clearance. The associated response would be to check rack access, airflow and maintenance clearances before the final deployment layout. It does not describe the central reason for this case, which is that fire detection and safety systems must be planned alongside accelerator and cooling deployment.
C:That principle is relevant to this distinct case: A new AI room has reliable cooling capacity but no monitoring for localized humidity or leak alarms. The associated response would be to add appropriate environmental sensors and integrate alerts with owned operational response procedures. It does not describe the central reason for this case, which is that fire detection and safety systems must be planned alongside accelerator and cooling deployment.
D:That principle is relevant to this distinct case: A new cluster has redundant high-speed links but cable trays cannot support the planned fiber and copper bundles. The associated response would be to design documented cable routes, bend radii and labeling for power and network infrastructure. It does not describe the central reason for this case, which is that fire detection and safety systems must be planned alongside accelerator and cooling deployment.