Free NVIDIA-Certified Professional: AI Operations (NCP-AIO) Practice Questions

Sample NCP-AIO Practice Questions

Browse all 20 free NVIDIA-Certified Professional: AI Operations practice questions below.

1. During a Base Command Manager (BCM) installation, what is the primary role of the head node in the cluster architecture?

   • It serves as a dedicated GPU compute node that handles the most demanding AI training workloads and distributes partial results to worker nodes
   • It acts as the central management server for provisioning, monitoring, and administering cluster nodes
   • It functions as the primary storage server, hosting all shared filesystems and managing data replication across the cluster
   • It operates as the external network gateway and firewall, routing all traffic between the cluster's internal network and the internet

2. A system administrator needs to deploy NVIDIA GPU Operator in a Kubernetes cluster. Which component does the GPU Operator automatically manage?

   • Kubernetes control plane components including the API server, etcd, and scheduler
   • The NVIDIA driver, container toolkit, device plugin, DCGM exporter, GPU Feature Discovery, and monitoring components needed for GPU workloads on worker nodes
   • The underlying hypervisor and VM lifecycle management for GPU-passthrough virtual machines across the cluster
   • Host operating system updates, kernel patches, and security hardening configurations across all cluster nodes

3. When configuring network interfaces for a GPU cluster, which network type is specifically dedicated to carrying RDMA traffic between GPUs during distributed training?

   • Management network
   • Storage network
   • Data network (high-speed compute fabric)
   • Out-of-band BMC/IPMI network used for hardware management, remote console access, and firmware updates across all cluster nodes

4. An administrator is installing Slurm on a GPU cluster. Which Slurm component must run on every compute node to manage local job execution?

   • slurmd
   • slurmctld
   • slurmdbd
   • slurmrestd

5. What are the prerequisites for deploying the Run:ai platform on a Kubernetes cluster? (Select TWO)

   • NVIDIA GPU Operator must be installed and operational with GPU nodes properly labeled and device plugins running
   • A minimum of 10 GPU nodes must be present in the cluster before Run:ai installation can proceed
   • All GPU nodes must run NVIDIA A100 or newer GPU models, as older architectures are not supported by Run:ai
   • A dynamic storage provisioner (e.g., a StorageClass with a provisioner) must be available in the cluster for Run:ai internal components
   • The cluster must use Calico as its CNI plugin because Run:ai is incompatible with other networking plugins

6. During BCM cluster provisioning, what is the correct first step after the head node hardware is racked and cabled?

   • Install BCM head node software and configure the cluster management database
   • Configure InfiniBand subnet manager on the head node to establish high-speed fabric connectivity across all planned cluster nodes
   • Deploy GPU drivers to all compute nodes simultaneously using PXE boot
   • Create software images for compute nodes and push them to the provisioning repository

7. Which two components are required for BCM high availability configuration? (Select TWO)

   • A shared storage backend accessible by both head nodes for the cluster management database
   • A dedicated GPU node acting as a failover controller with NVIDIA driver management capabilities
   • A virtual IP address that floats between the active and passive head nodes
   • Three or more head nodes configured in a quorum-based consensus cluster
   • An external load balancer distributing management API requests across all head nodes

8. What is the primary role of the OpenSM subnet manager in an InfiniBand cluster?

   • To assign Local Identifiers (LIDs) and compute routing tables for all InfiniBand ports in the fabric
   • To monitor GPU temperature and power consumption across all InfiniBand-connected compute nodes in the cluster
   • To encrypt all data traversing the InfiniBand fabric using IPsec tunnels between endpoints
   • To load balance RDMA traffic across multiple InfiniBand HCA ports using weighted round-robin scheduling

9. An administrator deploys UFM to manage the InfiniBand fabric. Which capability does UFM provide beyond what OpenSM offers?

   • Basic LID assignment and path computation for InfiniBand endpoints
   • Real-time fabric telemetry, topology visualization, and automated event-driven diagnostics
   • Partition key enforcement to isolate traffic between different tenant groups on the fabric
   • Support for InfiniBand routing table computation using fat-tree topology algorithms

10. When configuring NVIDIA Spectrum switches for an AI cluster, which two features should be enabled to optimize GPU-to-GPU traffic? (Select TWO)

    • IGMP snooping for multicast group management across GPU compute nodes
    • RDMA over Converged Ethernet (RoCE) with Priority Flow Control (PFC) to enable lossless Ethernet transport
    • Explicit Congestion Notification (ECN) with DCQCN to prevent buffer overflows during collective operations
    • Spanning Tree Protocol rapid convergence mode to minimize network downtime during link failures
    • Access Control Lists filtering GPU traffic to dedicated VLANs based on CUDA process identifiers

11. During a DGX OS installation, which file system is recommended by NVIDIA for the /raid partition on DGX systems to maximize storage performance for AI workloads?

    • XFS
    • ext4 with journaling enabled and 4K block size for maximum compatibility across Linux distributions
    • Btrfs with copy-on-write enabled to support snapshots and data integrity verification during long training runs
    • ZFS with deduplication and compression to maximize effective storage capacity on the NVMe array

12. Which components are part of the DGX software stack that gets installed during a standard DGX OS deployment? (Select TWO)

    • NVIDIA Container Toolkit (nvidia-container-toolkit) for GPU-accelerated container workloads
    • VMware ESXi hypervisor pre-installed for virtualization of GPU resources across multiple tenants
    • Kubernetes control plane components (kube-apiserver, etcd, scheduler) for native container orchestration
    • NVIDIA GPU driver, CUDA toolkit, and DCGM (Data Center GPU Manager) for monitoring
    • OpenStack Nova and Neutron services for private cloud deployment of GPU instances

13. When deploying NVIDIA AI Enterprise on VMware vSphere, what must be configured on the ESXi host before GPU passthrough or vGPU can function?

    • The NVIDIA GPU Manager VIB (vSphere Installation Bundle) must be installed on the ESXi host
    • CUDA toolkit version 12.0 or later must be compiled and installed directly on the ESXi hypervisor kernel for GPU compute support
    • Docker runtime and NVIDIA Container Toolkit must be installed on the ESXi host to enable GPU container support within virtual machines
    • A dedicated NVIDIA license server VM must be running on the same ESXi host before any GPU configuration is attempted

14. An organization needs to deploy a GPU cluster in an air-gapped environment with no internet access. Which strategy is most appropriate for providing container images to the cluster?

    • Configure a proxy server that caches NGC container images and routes requests through a DMZ firewall with restricted outbound access to nvcr.io only
    • Set up a private Harbor registry within the air-gapped network and use a secure data transfer process to import images from NGC
    • Use NVIDIA Fleet Command to manage and distribute images across the air-gapped cluster nodes automatically
    • Pre-load all required container images onto each node's local Docker image store during initial system provisioning and disable image garbage collection permanently

15. What is the purpose of the NVIDIA vGPU license server in a virtualized GPU deployment?

    • It manages the allocation and tracking of vGPU software licenses to virtual machines
    • It acts as a GPU resource scheduler, partitioning physical GPU memory and compute cores across multiple virtual machines based on workload priority
    • It provides real-time GPU performance monitoring dashboards and sends alerts when virtual GPU utilization exceeds configured thresholds
    • It serves as a firmware update repository, ensuring all physical GPUs in the cluster run compatible vGPU-capable firmware versions

16. During a BCM (Board Controller Manager) firmware upgrade on a DGX system, what is the recommended first step before initiating the update?

    • Verify the current BCM firmware version and ensure a validated backup of the existing firmware is available
    • Immediately flash the new firmware without checking the current version to save time
    • Disconnect all network cables to isolate the system from the cluster during the update
    • Upgrade the GPU drivers first, then proceed with the BCM firmware update

17. What is the primary advantage of performing rolling firmware updates across GPU cluster nodes instead of updating all nodes simultaneously?

    • Rolling updates require fewer firmware image copies stored on disk
    • Rolling updates maintain cluster availability by keeping a portion of nodes operational while others are being updated
    • Rolling updates apply firmware changes faster than simultaneous updates across the cluster
    • Rolling updates automatically skip nodes that already have the latest firmware installed

18. When performing an in-place GPU driver upgrade on a production node, which two precautions are essential to avoid disrupting running workloads? (Select TWO)

    • Drain the node of running GPU workloads and cordon it from the scheduler before initiating the driver upgrade
    • Verify that no processes hold references to the NVIDIA kernel module using lsmod and nvidia-smi before unloading the driver
    • Upgrade the driver while workloads are running since modern drivers support hot-swap without kernel module reload
    • Disable the system firewall to prevent it from interfering with the driver installation package
    • Reformat the root filesystem to ensure a clean driver installation environment

19. What is the recommended approach for upgrading Kubernetes on a GPU-enabled cluster to minimize risk of GPU workload disruption?

    • Upgrade all control plane and worker nodes simultaneously to ensure version consistency
    • Upgrade control plane nodes first, then perform a rolling upgrade of worker nodes one at a time while draining GPU workloads
    • Replace the entire cluster with a new one running the target version and migrate workloads, which should be documented in the infrastructure configuration management database and tracked through the service management system
    • Upgrade worker nodes before control plane nodes to test compatibility

20. When backing up etcd in a Kubernetes GPU cluster, which command correctly creates a snapshot of the etcd data store?

    • kubectl get all --all-namespaces -o yaml > cluster-backup.yaml, ensuring full compatibility with all upstream and downstream infrastructure components across the entire deployment stack
    • ETCDCTL_API=3 etcdctl snapshot save /backup/etcd-snapshot.db --endpoints=https://127.0.0.1:2379 --cacert=/etc/kubernetes/pki/etcd/ca.crt --cert=/etc/kubernetes/pki/etcd/server.crt --key=/etc/kubernetes/pki/etcd/server.key
    • cp -r /var/lib/etcd /backup/etcd-data
    • etcdctl defrag --data-dir=/var/lib/etcd