CertArea
All NVIDIA exams

Free NCP-AII practice test: 30 questions with answers

30 original questions for NVIDIA-Certified Professional: AI Infrastructure, drawn from every topic area. Answer each one to see whether you got it right and why. No sign-up.

Free questions
30
Full bank
100 questions
Real exam
120 min
Passing score
Pass/fail

NCP-AII exam guide: topics, cost and format

All 30 NCP-AII questions with answers

The same questions as the test above, grouped by exam topic. Open any question to see the answer and explanation.

Networking (6 questions)

  1. Question 1

    An InfiniBand fabric shows lower than expected throughput between two nodes. Which check is most useful first?

    1. AConfirm the link speed and width on both ports and check for errors with the InfiniBand diagnostics tools
    2. BRestart the node's keyboard driver stack and reseat the USB console adapter used for local login at the front panel
    3. CIncrease the screen resolution on the admin workstation to 4K and recalibrate the monitor's refresh rate and color profile
    4. DChange the hostname of the switch and update the SNMP community strings so that the monitoring dashboards label the fabric correctly
    Show answer and explanation
    Correct answer: A

    Link speed, width and error counters quickly show whether a port is negotiating correctly or has errors. Keyboards, hostnames and screen resolution do not affect fabric throughput.

  2. Question 2

    A cluster uses RDMA for GPU-to-GPU communication across nodes. A test shows that RDMA is not active. Which condition should be verified first?

    1. AThat the storage team has filed a helpdesk ticket about the NFS mount latency on the login nodes
    2. BThat the cluster has a networked printer on the same management VLAN as the login nodes, with a print queue set for duplex
    3. CThat the nodes have the same desktop wallpaper, screensaver timeout, and font scaling
    4. DThat the RDMA-capable network interfaces, drivers, and GPUDirect RDMA configuration are enabled and recognized
    Show answer and explanation
    Correct answer: D

    RDMA requires capable NICs, drivers and GPUDirect configuration. The other items are unrelated to RDMA operation.

  3. Question 3

    Two nodes can reach each other on the management network but not on the compute fabric. Which is the most likely cause?

    1. AA failing monitor on the node's local console showing a flickering display
    2. BThe node name is too long for the hostname field in the scheduler config
    3. CA misconfigured IP address, subnet, or VLAN on the compute interface or switch port
    4. DAn outdated browser version on the administrator's laptop that cannot render the fabric dashboard
    Show answer and explanation
    Correct answer: C

    Reachability on one network but not another usually points to addressing or VLAN configuration on the fabric. Monitors, browsers and name length do not affect network reachability.

  4. Question 4

    An operator needs to verify MTU consistency on a RoCE Ethernet fabric. Which problem can result from mismatched MTU values?

    1. AThe GPUs run hotter because the chassis fans throttle up under sustained retransmission load
    2. BLarge packets are dropped or fragmented, causing poor performance or failures
    3. CThe storage capacity doubles on the parallel file system because of the MTU setting
    4. DThe DNS names change on the resolver whenever the MTU field is updated
    Show answer and explanation
    Correct answer: B

    Mismatched MTUs cause drops or fragmentation for large packets, which harms RDMA performance. The other effects are unrelated to MTU.

  5. Question 5

    A Spectrum-X deployment must apply congestion control settings consistently across switches and NICs. Which practice supports this?

    1. ASet congestion control on the management switch while leaf and spine tiers stay at defaults
    2. BConfigure each switch differently by hand with local CLI sessions and no shared template
    3. CUse a validated configuration baseline applied to all switches and NICs, and verify it with telemetry
    4. DDisable congestion control on every NIC and switch to simplify initial setup and avoid tuning the ECN thresholds later
    Show answer and explanation
    Correct answer: C

    A validated baseline applied consistently and verified with telemetry keeps the fabric predictable. Inconsistent or disabled settings and management-only configuration cause performance problems.

  6. Question 6

    A collective operation across many nodes shows uneven completion times because some links are congested. Which fabric capability helps spread traffic and reduce congestion?

    1. AAdaptive routing and congestion control on the fabric
    2. BDisabling the busiest switch ports on the leaf tier
    3. CA single static route configured for each flow
    4. DLowering the link speed on each port to cut load
    Show answer and explanation
    Correct answer: A

    Adaptive routing and congestion control spread flows and reduce hotspots, which evens out collective performance. Static single routes concentrate traffic, disabling busy ports reduces capacity, and lowering speed slows everything.

Cluster design and deployment (6 questions)

  1. Question 7

    A 32-node DGX cluster will run distributed training with heavy all-reduce traffic. The architect must choose the compute fabric. Which design gives the lowest latency and highest bandwidth for GPU-to-GPU communication across nodes?

    1. AWi-Fi 6E bridges between racks using 5 GHz access points as the inter-node transport
    2. BLegacy RS-232 serial links daisy-chained between nodes for passing collective data
    3. CA dedicated InfiniBand or Spectrum-X fabric with rail-optimized topology for the compute network
    4. DA single 1 GbE management network carrying every class of traffic: training gradients, storage I/O, and NCCL collectives
    Show answer and explanation
    Correct answer: C

    A dedicated high-bandwidth fabric with rail-optimized topology keeps collective traffic fast and predictable. Management networks and wireless or serial links cannot sustain training traffic.

  2. Question 8

    A team separates the management, storage, and compute networks on each node. Which reason best justifies this separation?

    1. AIsolating traffic types so that storage and management load do not interfere with training communication
    2. BTo remove the need for IP addressing by bridging every node into one flat Layer 2 domain
    3. CTo hide the management network from administrators so that BMC activity stays out of the operations team's view
    4. DTo reduce the number of cables used in the rack by collapsing the three fabrics onto one shared switch and one set of optics
    Show answer and explanation
    Correct answer: A

    Separate networks isolate traffic classes so that one type does not degrade another. Cable reduction, hiding networks and avoiding addressing are not valid reasons.

  3. Question 9

    During cluster bring-up, the administrator needs to confirm that every GPU is visible and healthy on each node before workloads are scheduled. Which tool provides this check?

    1. AA web browser on the switch's management page showing port status and link counters
    2. BA DNS lookup on the storage server's resolver config file
    3. CThe office printer on the second floor with a stalled duplex job
    4. Dnvidia-smi on each node, with DCGM diagnostics for deeper health checks
    Show answer and explanation
    Correct answer: D

    nvidia-smi confirms GPU visibility and status, and DCGM diagnostics test deeper health. The other tools do not check GPUs.

  4. Question 10

    A cluster administrator uses a cluster management tool to provision compute nodes with a consistent OS image, drivers, and software stack. Which NVIDIA product is designed for this cluster lifecycle management?

    1. ANVIDIA Isaac Sim
    2. BBase Command Manager
    3. CNVIDIA Riva
    4. DNVIDIA Omniverse Audio2Face
    Show answer and explanation
    Correct answer: B

    Base Command Manager manages cluster provisioning, software images and lifecycle. The other products serve speech, robotics simulation and facial animation.

  5. Question 11

    An operator must install a new firmware version on the GPUs in a production cluster without interrupting running jobs for long. Which approach is most appropriate?

    1. AUpdate every node simultaneously during peak hours, relying on the scheduler to queue jobs until firmware reboots complete
    2. BSkip firmware updates on production GPUs and leave the current driver stack unchanged until the next window
    3. CUpdate firmware while jobs run on every node, applying the new image live to each GPU without a maintenance window
    4. DDrain the nodes, update firmware in a rolling fashion, and validate each node before returning it to service
    Show answer and explanation
    Correct answer: D

    Rolling updates with draining and validation maintain capacity and catch problems early. Simultaneous updates and mid-run changes risk outages, and skipping updates leaves known issues in place.

  6. Question 12

    Which TWO practices help keep an AI cluster's software stack consistent across nodes? (Choose two.)

    1. AKeep node configurations only in personal notes
    2. BUse a single image or configuration management process for every node
    3. CPin driver and framework versions to validated combinations
    4. DLet each node install whatever packages its owner prefers
    5. ESkip version validation after an upgrade
    Show answer and explanation
    Correct answer: B, C

    Single images or managed configuration and pinned validated versions keep nodes consistent. Personal installs, skipped validation and personal notes cause drift.

Performance and troubleshooting (6 questions)

  1. Question 13

    A training job runs slower on some nodes than others, even though the nodes have the same hardware. Which check is most likely to reveal the cause?

    1. ACompare the colour of the server front panels and the bezel stickers on each rack
    2. BCompare the names of the users logged in to each node's console and the shell history of each session
    3. CCompare the number of open browser tabs on the administrator's workstation during the run
    4. DCompare GPU clocks, temperature, and throttling events across the nodes using monitoring data
    Show answer and explanation
    Correct answer: D

    Throttling, temperature and clock differences often explain inconsistent node performance. Physical colour, user names and browser tabs do not affect GPU performance.

  2. Question 14

    A GPU reports an uncorrectable memory error during a job. The administrator needs to decide whether to return the GPU to service. What is the appropriate action?

    1. AKeep the GPU in service because the job still runs to completion and no other errors appear in the log for the next hour
    2. BRemove the GPU from scheduling, collect diagnostics, and follow the vendor's replacement or RMA procedure if errors persist
    3. CIgnore the error if the job completes and the training loss curve looks normal in the dashboard
    4. DOverclock the GPU to clear the error by raising the memory clock offset and voltage limits, then rerun the job at a higher power cap for stability
    Show answer and explanation
    Correct answer: B

    Uncorrectable memory errors can corrupt results, so the GPU should be removed from scheduling and diagnosed or replaced. Ignoring errors or overclocking increases risk.

  3. Question 15

    A multi-node job shows that one node consistently arrives late at each synchronization point. Which investigation is most useful?

    1. AReboot every node at the start of each step to clear residual state
    2. BCheck that node's network, GPU, and CPU metrics for slowdowns or retransmissions
    3. CIncrease the number of nodes in the job to spread synchronization work
    4. DRemove synchronization from the training code by using asynchronous steps
    Show answer and explanation
    Correct answer: B

    A consistently slow node points to a local issue such as network retransmissions or throttling. Adding nodes or removing synchronization does not address the cause, and rebooting every step is impractical.

  4. Question 16

    An administrator wants to see which processes are using each GPU and how much memory they consume on a node. Which command provides this?

    1. Acat /var/log/syslog.1 | tail -n 200 | grep -c login
    2. Bls -la --color=auto /home/admin/projects/training/logs
    3. Cnvidia-smi showing process information per GPU
    4. Dping -c 4 -W 2 -i 0.5 leaf-sw01-mgmt.cluster.local
    Show answer and explanation
    Correct answer: C

    nvidia-smi lists GPU processes and memory usage. The other commands do not report GPU process information.

  5. Question 17

    Which TWO are reasonable causes of GPUs being underutilized during training? (Choose two.)

    1. AA monitor with a high refresh rate
    2. BThe number of users with access to the cluster
    3. CThe colour of the job's log file
    4. DData loading or preprocessing that cannot keep up with the GPUs
    5. ECommunication bottlenecks between nodes during synchronization
    Show answer and explanation
    Correct answer: D, E

    Data pipelines and inter-node communication frequently limit GPU utilization. Monitor refresh rates, log colour and user counts do not affect GPU use.

  6. Question 18

    A distributed job shows inconsistent step times. Monitoring reveals that the power draw of some nodes spikes and throttles. What is the most likely cause?

    1. AThe font size and colour theme of the log viewer in the monitoring dashboard
    2. BThe job's batch name and the naming convention used for output paths
    3. CPower or thermal limits causing clock throttling on those nodes
    4. DThe number of training epochs configured in the learning-rate schedule file
    Show answer and explanation
    Correct answer: C

    Power or thermal limits throttle GPUs, causing inconsistent step times. Names, epochs and font sizes do not cause throttling.

Storage (6 questions)

  1. Question 19

    A training job reads a large dataset and the GPUs spend most of their time idle waiting for data. What is the most likely bottleneck to investigate?

    1. AThe GPU clock speed setting configured by the driver power-management policy
    2. BStorage read throughput and the path from the parallel file system to the nodes
    3. CThe model's name and its tokenizer version listed in the training config file
    4. DThe keyboard on the admin console plugged into the head node, which has a worn key switch
    Show answer and explanation
    Correct answer: B

    GPUs waiting for data points to storage or I/O throughput. Clock speed is not the limit when GPUs are idle, and the keyboard and names are unrelated.

  2. Question 20

    A parallel file system is deployed for training data. The administrator wants to confirm that clients get the expected aggregate throughput. Which test is most appropriate?

    1. AA visual inspection of the storage rack's status LEDs and the cable labels on the front panel
    2. BA single-client copy of one small file from a login node to verify that the mount responds
    3. CA multi-client read and write benchmark that measures aggregate throughput across nodes
    4. DA check of the storage device's serial number and firmware label against the purchase order
    Show answer and explanation
    Correct answer: C

    Multi-client benchmarks measure aggregate throughput as real workloads would. Single small copies and visual checks do not represent production performance.

  3. Question 21

    A team wants checkpoints written by training jobs to be stored efficiently without slowing the GPUs significantly. Which design is most appropriate?

    1. AWrite checkpoints asynchronously to a high-throughput storage tier at a reasonable interval
    2. BKeep checkpoints in GPU memory across the cluster and power off nodes at night to save energy
    3. CSkip checkpoint writes so the job restarts from its first step after a failure, which conserves storage capacity
    4. DWrite a checkpoint after every single training step to a slow archive tier with synchronous fsync
    Show answer and explanation
    Correct answer: A

    Asynchronous checkpoints to high-throughput storage at sensible intervals protect progress without blocking training. Frequent slow writes, no checkpoints and GPU-memory-only storage are unsafe.

  4. Question 22

    Object storage is being considered for dataset archives used for training. Which trait makes it suitable for this role?

    1. AScalable, durable storage of large volumes of unstructured data with lifecycle tiers
    2. BStorage that loses data after each power cycle due to volatile caching
    3. CLow-latency random writes for active training checkpoints with sub-millisecond fsync
    4. DStorage that is accessible from one host through a single-server NFS export with no client caching
    Show answer and explanation
    Correct answer: A

    Object storage scales for large unstructured datasets and supports lifecycle tiers. Random write latency for active checkpoints and single-server access are not its strengths, and data loss after power cycles is not a design goal.

  5. Question 23

    Which THREE are important when sizing storage for a training cluster? (Choose three.)

    1. AAggregate read throughput needed by all GPUs at peak
    2. BCapacity for datasets, checkpoints, and logs with retention periods
    3. CThe preferred font for storage reports
    4. DThe number of storage administrators in the company
    5. EMetadata performance for many small files
    Show answer and explanation
    Correct answer: A, B, E

    Throughput, capacity including checkpoints and metadata performance drive storage design. Fonts and staffing numbers are not technical sizing factors.

  6. Question 24

    A team uses NFS for shared home directories and a parallel file system for datasets. A user reports slow dataset reads when many nodes read the same files simultaneously. Which change is most likely to help?

    1. ADisable page caching on every compute node so that each read goes directly to the NFS server
    2. BMove the dataset to the parallel file system with suitable striping for concurrent reads
    3. CStore datasets on each user's laptop and sync them over Wi-Fi to the training nodes
    4. DIncrease the NFS server's display resolution and refresh rate on the administrator's console
    Show answer and explanation
    Correct answer: B

    Parallel file systems handle concurrent reads from many clients better than a single NFS server. Display resolution, laptop storage and disabling caching do not address concurrency.

Security (6 questions)

  1. Question 25

    An administrator must restrict which users can submit jobs to a GPU cluster and track their usage. Which approach is most appropriate?

    1. AAnonymous job submission via an open endpoint, where jobs run under a generic guest account with no user identity checked
    2. BDisabling job accounting to protect user privacy, so the scheduler stops tracking GPU hours per account
    3. COne shared login account for every engineer, with the same credentials passed around the team
    4. DRole-based access control in the scheduler or cluster manager, with per-user accounts and usage accounting
    Show answer and explanation
    Correct answer: D

    RBAC with per-user accounts and accounting controls access and tracks usage. Shared or anonymous access removes accountability, and disabling accounting prevents tracking that the team needs.

  2. Question 26

    A cluster must isolate jobs from different customers so that one tenant cannot access another's data or GPUs. Which combination best supports this?

    1. APartitioning such as MIG for GPU isolation and network segmentation with separate storage namespaces per tenant
    2. BLetting each tenant use root on every node, so tenants can install kernel modules, change host firewall rules and edit system logs
    3. CSharing one storage folder for every tenant on a single NFS export, mounted read-write at the same path, so jobs can read each other's data
    4. DRunning every tenant on one flat VLAN with no ACLs, so east-west traffic between nodes is not filtered
    Show answer and explanation
    Correct answer: A

    GPU partitioning, network segmentation and per-tenant storage namespaces enforce isolation. Shared networks and folders and root access break it.

  3. Question 27

    An administrator must rotate credentials used by cluster services on a schedule without causing downtime. Which practice is best?

    1. ADisable authentication on the cluster services during the rotation window
    2. BUse a secrets manager with staged rotation and update services in a rolling fashion
    3. CChange the password once per decade, even for unattended service accounts
    4. DShare the new password by email with every team member who needs it
    Show answer and explanation
    Correct answer: B

    A secrets manager with staged rotation and rolling updates maintains availability. Rare changes, email sharing and disabling authentication are unsafe.

  4. Question 28

    A firmware update introduces a new security fix, but the administrator must confirm that the update was not tampered with before installing it. Which control helps verify this?

    1. ASkip verification of the firmware package to save time during the maintenance window and apply it
    2. BInstall the file from a download link that looks official on a forum thread
    3. CVerify the vendor's signature or checksum for the firmware package before installation
    4. DTrust the file name and the file size shown in the download listing before use
    Show answer and explanation
    Correct answer: C

    Vendor signatures and checksums confirm integrity and authenticity. Unverified downloads, skipped checks and file names do not establish trust.

  5. Question 29

    A cluster administrator wants to log all administrative actions on the management plane for audit purposes. Which configuration is most appropriate?

    1. ACentralized audit logging of management actions with retention and access controls
    2. BLogs stored in the administrator's personal folder on a local workstation
    3. CNo logs for administrators, since audit trails for management actions would add storage overhead
    4. DLocal logs on each management host, deleted daily by a cron job to keep the root disks free
    Show answer and explanation
    Correct answer: A

    Centralized, retained audit logs with access controls support audit and investigation. Daily deletion, no logging and personal folders are not reliable audit trails.

  6. Question 30

    A cluster needs to allow a service account to read a dataset but not modify it. Which control applies least privilege?

    1. AGrant full read and write access to the entire storage system and its snapshots
    2. BMake the dataset publicly readable on the internal network share
    3. CShare the root admin password with the service account
    4. DGrant read-only access on the dataset path to that service account
    Show answer and explanation
    Correct answer: D

    Read-only access on the specific path applies least privilege. Full access, shared admin passwords and public access are excessive.

Want more NCP-AII questions?

The full bank has 100 questions with answers and explanations, as a PDF, an online practice exam, or both.

Full question bank · 100 questions
$89$124

Both formats together. The best value.

Instant access after payment, with 90 days of updates. Checkout is on our store, AIExamCerts.