NVIDIA-Certified Professional: AI Infrastructure (NCP-AII)

Exam practice questions

15 / 176 questions

Free preview · Answers and explanations includedPage 1 of 2

Troubleshooting and OptimizationNCP-AII practice

A DGX power alert identifies one supply bay. Which investigation should precede concluding that the supply module itself is defective?

Select one answerAReplace the supply before checking whether its input feed is present.BCompare GPU utilization with a different server’s utilization.CReset all power alerts and accept the bay if the host stays online.DCheck the identified bay, BMC status, feed, and connection.

Control Plane Installation and ConfigurationNCP-AII practice

A Pyxis job fails because a required dataset path is absent inside the container, although the host mount is healthy. What should the administrator check next?

Select one answerARemount the already working host filesystem.BCheck container mounts and job-user access.CIncrease GPU count without adding dataset mounts.DRebuild the image with unchanged bind mounts.

Physical Layer ManagementNCP-AII practice

A BlueField in DPU mode exposes a host network function, but host traffic does not pass after boot. The Arm-side virtual switch has not loaded. Which dependency should be addressed first?

Select one answerAReinstall the host driver while leaving DPU switching unavailable.BTreat host interface enumeration as proof that DPU forwarding is ready.CRestore the required Arm-side virtual-switch configuration.DChange only host routes without checking the missing DPU-side dependency.

Systems and Server Bring-UpNCP-AII practice

A technician records GPU index 0 as the identity of a physical accelerator. After maintenance, enumeration order changes. Which identifier should the inventory use to track the same GPU more reliably?

Select one answerAThe current device index without a UUID.BThe PID of the last GPU workload.CAn earlier container-local GPU ordinal.DGPU UUID associated with the current asset record.

Systems and Server Bring-UpNCP-AII practice

An administrator is commissioning out-of-band management for a GPU rack. Which TWO checks demonstrate that the configuration supports later troubleshooting? Select TWO.

Select two answersATest authorized access to the correct BMC endpoints.BValidate host SSH access as the only proof of BMC recovery readiness.CAssociate inventory and event times with the physical servers.DRetain BMC events without mapping their identifiers or times to host records.

Systems and Server Bring-UpNCP-AII practice

A BMC is reachable by IP address but its recorded event times are several hours different from host logs. The operations team must correlate intermittent power events. What should be corrected first?

Select one answerACorrelate by log order despite clock offsets.BClear logs and wait for another incident.CAlign and verify management and host time sources.DAdjust timeouts without reconciling event clocks.

Cluster Test and VerificationNCP-AII practice

A standard NVIDIA GPU HPL run will use four nodes with eight GPUs each and one GPU per MPI process. Which TWO launch/input conditions are consistent with that plan? Select TWO.

Select two answersALaunch 32 MPI processes with one process placed on each GPU.BSet a valid HPL process grid with P multiplied by Q equal to 32.CLaunch eight MPI processes with one process assigned to each node pair.DSet P multiplied by Q equal to four to match the node count.

Cluster Test and VerificationNCP-AII practice

An engineer will run an intrusive GPU diagnostic on a node. Which TWO preparations make its results meaningful and avoid contention with user jobs? Select TWO.

Select two answersAKeep a representative training job active to provide a realistic baseline.BReserve the node and meet the test’s exclusivity requirements.CReset event history before recording the initial diagnostic state.DRecord initial health and retain results with correlated telemetry.

Control Plane Installation and ConfigurationNCP-AII practice

Slurm advertises eight GPUs on a node that currently exposes only seven usable devices. New GPU jobs are still eligible to land there. What is the safest operational response while the mismatch is investigated?

Select one answerAIncrease advertised count to offset missing hardware.BRestrict new allocations and reconcile the resource mismatch.CAllocate the eighth GPU without verifying it.DSuppress warnings and retain the advertised count.

Systems and Server Bring-UpNCP-AII practice

The cluster inventory lists the expected GPUs, but a topology check shows an adapter attached through a different CPU socket from the approved layout. Why should the discrepancy be investigated before performance acceptance?

Select one answerAEqual device counts establish equivalent data paths.BThe changed attachment can alter locality and supported paths.CAttachment changes affect management traffic only.DKeep old CPU affinity without checking topology.

Troubleshooting and OptimizationNCP-AII practice

One GPU reports reduced clocks during a sustained run. Its power telemetry shows that it is operating at the configured power limit, while temperatures remain within range. What should be investigated before declaring a hardware fault?

Select one answerAA thermal failure solely because the clocks are lower than peak specification.BWhether the configured power limit and expected workload performance match the approved system policy.CA missing GPU solely because utilization remains nonzero.DA required driver removal without reviewing the power configuration.

Cluster Test and VerificationNCP-AII practice

An NCCL test is intended to measure traffic between two racks. Which TWO checks establish that its scope actually crosses the rack boundary? Select TWO.

Select two answersARecord rank-to-node and node-to-rack placement.BVerify cross-rack group membership and the intended fabric traffic.CInfer cross-rack communication from the job’s GPU count alone.DInfer physical scope from the split operator without rank placement.

Systems and Server Bring-UpNCP-AII practice

A DGX system reports a failed power-supply input during commissioning, but all GPUs are currently visible. The acceptance requirement includes power redundancy. How should the result be handled?

Select one answerAApprove redundancy because GPUs remain visible.BResolve the failed input before accepting redundancy.CUse reduced idle draw as redundancy evidence.DClear the alert while the OS is reachable.

Control Plane Installation and ConfigurationNCP-AII practice

An operating-system image update includes a new kernel. The GPU kernel module was built only for the previous kernel. What must be verified before that image is deployed broadly?

Select one answerACheck that the old module file exists.BVerify the new kernel’s supported module loads after boot.CApprove from package installation status alone.DTest containers without validating the host module.

Cluster Test and VerificationNCP-AII practice

A single server has GPU ranks 0 through 7. NCCL_TESTS_SPLIT="MOD 2" is set. How are the ranks grouped?

Select one answerAEight groups: one rank in each group.BOne group: all eight ranks together.CTwo groups: even ranks and odd ranks.DTwo groups: ranks 0–3 and ranks 4–7.

Complete your exam preparation

Continue with the full question bank

Practise free with answers, explanations, and Community discussions. Join Community to unlock the complete question bank and timed mock exams.

Join Community Choose access for your certification

NVIDIA-Certified Professional: AI Infrastructure (NCP-AII)

Exam practice questions

15 / 176 questions

Free preview · Answers and explanations includedPage 1 of 2

Troubleshooting and OptimizationNCP-AII practice

A DGX power alert identifies one supply bay. Which investigation should precede concluding that the supply module itself is defective?

Control Plane Installation and ConfigurationNCP-AII practice

A Pyxis job fails because a required dataset path is absent inside the container, although the host mount is healthy. What should the administrator check next?

Physical Layer ManagementNCP-AII practice

A BlueField in DPU mode exposes a host network function, but host traffic does not pass after boot. The Arm-side virtual switch has not loaded. Which dependency should be addressed first?

Systems and Server Bring-UpNCP-AII practice

A technician records GPU index 0 as the identity of a physical accelerator. After maintenance, enumeration order changes. Which identifier should the inventory use to track the same GPU more reliably?

Select one answerAThe current device index without a UUID.BThe PID of the last GPU workload.CAn earlier container-local GPU ordinal.DGPU UUID associated with the current asset record.

Systems and Server Bring-UpNCP-AII practice

An administrator is commissioning out-of-band management for a GPU rack. Which TWO checks demonstrate that the configuration supports later troubleshooting? Select TWO.

Systems and Server Bring-UpNCP-AII practice

A BMC is reachable by IP address but its recorded event times are several hours different from host logs. The operations team must correlate intermittent power events. What should be corrected first?

Cluster Test and VerificationNCP-AII practice

A standard NVIDIA GPU HPL run will use four nodes with eight GPUs each and one GPU per MPI process. Which TWO launch/input conditions are consistent with that plan? Select TWO.

Cluster Test and VerificationNCP-AII practice

An engineer will run an intrusive GPU diagnostic on a node. Which TWO preparations make its results meaningful and avoid contention with user jobs? Select TWO.

Control Plane Installation and ConfigurationNCP-AII practice

Slurm advertises eight GPUs on a node that currently exposes only seven usable devices. New GPU jobs are still eligible to land there. What is the safest operational response while the mismatch is investigated?

Systems and Server Bring-UpNCP-AII practice

The cluster inventory lists the expected GPUs, but a topology check shows an adapter attached through a different CPU socket from the approved layout. Why should the discrepancy be investigated before performance acceptance?

Troubleshooting and OptimizationNCP-AII practice

One GPU reports reduced clocks during a sustained run. Its power telemetry shows that it is operating at the configured power limit, while temperatures remain within range. What should be investigated before declaring a hardware fault?

Cluster Test and VerificationNCP-AII practice

An NCCL test is intended to measure traffic between two racks. Which TWO checks establish that its scope actually crosses the rack boundary? Select TWO.

Systems and Server Bring-UpNCP-AII practice

A DGX system reports a failed power-supply input during commissioning, but all GPUs are currently visible. The acceptance requirement includes power redundancy. How should the result be handled?

Control Plane Installation and ConfigurationNCP-AII practice

An operating-system image update includes a new kernel. The GPU kernel module was built only for the previous kernel. What must be verified before that image is deployed broadly?

Cluster Test and VerificationNCP-AII practice

A single server has GPU ranks 0 through 7. NCCL_TESTS_SPLIT="MOD 2" is set. How are the ranks grouped?

Select one answerAEight groups: one rank in each group.BOne group: all eight ranks together.CTwo groups: even ranks and odd ranks.DTwo groups: ranks 0–3 and ranks 4–7.

Complete your exam preparation

Continue with the full question bank

Practise free with answers, explanations, and Community discussions. Join Community to unlock the complete question bank and timed mock exams.

Join Community Choose access for your certification