NVIDIA Certified Professional - AI Infrastructure practice questions

Question 1

A developer is running the DGX H100/H200 quickstart health-check workflow before production and wants to know what the documented workflow covers beyond the basic `nvsm show health` step. According to the quickstart, which TWO are part of this bring-up workflow? (Select TWO.)

Accepted Answer

Verifying that the Docker installation is present on the system

Accepted Answer

Verifying that the NVIDIA driver installation is present on the system

Answer

Reconfiguring the NVSwitch fabric routing tables on each GPU board

Answer

Re-running the cluster-level NCCL AllReduce test across all nodes

Question 2

An architect is designing rack power for a DGX SuperPOD where multi-node jobs are not checkpointed and a single power-source failure must not drop a system. According to the Electrical Specifications, which provisioning scheme meets this availability requirement?

Accepted Answer

Three rPDUs per rack, each feeding two of six supplies, leaving four live

Answer

Two rPDUs per rack, each feeding three of the six power supplies on a system

Answer

Six rPDUs per rack, each feeding one supply, so any loss drops one PSU

Answer

One rPDU per rack feeding all six supplies from a single upstream path

Question 3

A sysadmin is initializing a single control-plane kubeadm cluster today but plans to upgrade it to high availability later. According to the kubeadm documentation, which step should be taken at init time to enable that future upgrade?

Accepted Answer

Specify --control-plane-endpoint at init to set a shared endpoint, such as a DNS name or load-balancer IP address.

Answer

Co-locate etcd on a separate baseboard now so the stacked topology can later be split without re-running kubeadm init.

Answer

Pass --apiserver-advertise-address pointing at the node's primary NIC so extra control-plane nodes can join later.

Answer

Set --pod-network-cidr to a reserved range so the load balancer can route traffic to future control-plane replicas.

Question 4

A platform team is rolling out distributed training across many nodes and wants NCCL debug logs written to per-process files. The team sets NCCL_DEBUG_FILE to a single static filename and finds logs are lost or corrupted. Given the constraints, how should the team fix this?

Accepted Answer

Use the filename.%h.%p pattern so %h and %p expand to a unique file per host and PID

Answer

Point NCCL_DEBUG_FILE at /dev/stdout so all processes append to one shared stream safely

Answer

Lower NCCL_DEBUG to WARN so fewer lines are written and the file stops corrupting

Answer

Set NCCL_DEBUG_SUBSYS to ALL so each subsystem writes to a separate debug log file

Question 5

A team is rolling out a new InfiniBand fabric and the operator needs continuous port telemetry — network bandwidth, congestion, errors, and latency — for the WebUI dashboards. According to the UFM Telemetry documentation, what is the default UFM Telemetry session interval used for monitoring and threshold events?

Accepted Answer

30 seconds, by default.

Answer

1 second, matching DCGM defaults.

Answer

5 minutes, to minimize fabric load.

Answer

1 hour, with manual refresh required for live data.

Question 6

An MLOps team is being deployed to run the NVIDIA HPL benchmark and needs to validate a matrix larger than aggregate GPU memory before production. Which setting activates HPL out-of-core mode so oversized matrix data spills to host CPU memory?

Accepted Answer

Set HPL_OOC_MODE=1 and specify a larger matrix size.

Answer

Set HPL_OOC_MODE=0 and reduce the matrix N parameter.

Answer

Set HPL_IN_CORE=1 to keep the matrix in GPU memory.

Answer

Set CUDA_VISIBLE_DEVICES to all GPUs on the host.

Question 7

A platform team is rolling out a fleet inventory tool that needs to know which components live on which tray of a DGX H100/H200 chassis. According to the DGX firmware update guide, which components live on the GPU tray versus the motherboard tray?

Accepted Answer

Motherboard tray: CPUs, PCH, BMC. GPU tray: GPUs, NVSwitches, HMC.

Answer

Motherboard tray: BMC only. GPU tray: GPUs and every fabric ASIC.

Answer

Motherboard tray: GPUs, BMC, HMC. GPU tray: CPUs, NVSwitches, PCH.

Answer

Motherboard tray: NVSwitches, PCH, HMC. GPU tray: GPUs, BMC, CPUs.

Question 8

A team needs to choose between three Helm-driven feature flags on the GPU Operator chart for their new cluster: pinning a specific driver image, exposing GPU metrics, and configuring MIG. According to the GPU Operator getting-started documentation, which set of Helm options are part of the installer configuration surface? (Select TWO.)

Accepted Answer

driver.enabled for driver container deployment.

Accepted Answer

mig.strategy for Multi-Instance GPU configuration.

Answer

kubeadm.bootstrap=true to provision the control plane itself.

Answer

operator.disableObservability=true to remove all metrics endpoints.

Question 9

An administrator needs to run a quick first-pass health check on a DGX H100 system before deployment to confirm the components and driver are healthy. According to the DGX user guide, which command provides this system health summary?

Accepted Answer

Run sudo nvsm show health and confirm the summary reports all checks as Healthy

Answer

Run dcgmi diag -r 4 to execute the longest DCGM hardware diagnostic across the GPUs

Answer

Run nvidia-smi -q to dump the full per-GPU query and read the overall health field

Answer

Run nv-hostengine --version to confirm the management daemon matches the driver

Question 10

A network team is being deployed to validate cabling on a new InfiniBand fabric. The lead must decide which UFM tab to use to confirm that there is one and only one active master Subnet Manager. According to the UFM Fabric Health Tab documentation, which diagnostic does the system perform?

Accepted Answer

Verifies there is one and only one active (master) Subnet Manager in the fabric and evaluates link width and link speed usage.

Answer

Runs continuous DCGM Diag suites against every connected GPU to identify cables that are likely failing.

Answer

Issues NCCL all-reduce tests across each fabric link to score effective throughput between pairs of endpoints.

Answer

Performs an automated Subnet Manager re-election every five minutes to keep routing tables freshly converged.

Question 11

A production team is designing its operational runbook around Mission Control autonomous hardware recovery for GB200 trays. Which TWO behaviors does NVIDIA document as part of the automated break/fix workflow? (Select TWO.)

Accepted Answer

It runs comprehensive validation to ensure systems are properly restored to service.

Accepted Answer

It executes a series of diagnostic steps to determine the cause of the failure.

Answer

It migrates running jobs to spare nodes using Slurm preemption before any repair.

Answer

It reflashes the GPU VBIOS on every tray as a mandatory precautionary first step.

Question 12

A platform team is being deployed to operationalize MIG on Hopper GPUs and must document exactly when a configured MIG mode will survive. Which TWO statements about MIG mode persistence are correct? (Select TWO.)

Accepted Answer

MIG mode is not persistent across system reboots because there is no longer a status bit in the GPU InfoROM

Accepted Answer

Unloading and reloading the driver kernel modules disables MIG mode because it persists only while loaded

Answer

MIG mode is written to the GPU InfoROM, so it survives both system reboots and full driver unloads

Answer

MIG mode is stored in the GPU's BMC and is restored automatically by Redfish on the next system cold boot

Question 13

An MLOps team is rolling out GPU sharing on a cluster of oversubscribed A100 GPUs and must explain to security reviewers how time-slicing differs from Multi-Instance GPU (MIG). According to the GPU Operator time-slicing documentation, which statement is accurate?

Accepted Answer

Time-slicing hands out GPU replicas to pods but gives no memory or fault isolation between replicas.

Answer

Time-slicing partitions each GPU into instances that receive isolated memory and fault domains per pod.

Answer

Time-slicing requires MIG to be disabled and then assigns one entire GPU exclusively to each pod.

Answer

Time-slicing depends on the cgroup driver to enforce strict per-pod GPU memory limits at runtime.

Question 14

An organization is migrating GPU telemetry to a containerized DCGM Exporter, and an engineer must avoid protocol incompatibilities between the exporter's DCGM client library and the host nv-hostengine. According to the documentation, which versioning rule should the engineer follow?

Accepted Answer

Use a DCGM version for the exporter that is greater than or equal to the host nv-hostengine

Answer

Pin the exporter's DCGM to an older release than the host so the host stays backward compatible

Answer

Match the exporter's DCGM exactly to the host CUDA driver branch instead of the nv-hostengine build

Answer

Run nvidia-fabricmanager on the host so the exporter can negotiate any DCGM version itself

Question 15

A production team is monitoring link quality and wants the system to raise an event automatically when a link's signal degrades, so operators can act before packet loss climbs. According to the Signal Degradation Monitoring documentation, how does the system detect a degraded link?

Accepted Answer

It continuously watches the link BER and generates an event when configured BER limits are crossed

Answer

It mirrors all traffic on the suspect port to a remote collector appliance for offline packet analysis

Answer

It disables FEC on the affected port so the operator can directly observe the raw uncorrected bit errors

Answer

It polls the transceiver temperature reading and flags any link that exceeds a thermal threshold

NVIDIA Certified Professional - AI Infrastructure

Sample questions

Sources

Similar exams