Free NVIDIA-Certified Professional: AI Networking (NCP-AIN) Practice Questions

Sample NCP-AIN Practice Questions

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

1. When configuring RoCE v2 on an NVIDIA Spectrum switch, which protocol encapsulation is used to transport RDMA traffic over an Ethernet network?

   • RDMA frames are encapsulated within GRE tunnels to provide isolation and routing across Layer 3 boundaries
   • UDP/IP encapsulation, where RDMA payloads are carried inside UDP datagrams with destination port 4791
   • Native InfiniBand frames carried directly on Ethernet using EtherType 0x8915 without any IP header
   • TCP/IP encapsulation with a dedicated RDMA stream identifier in the TCP options field for reliable delivery

2. What is the primary purpose of enabling Priority Flow Control (PFC) on a network configured for RoCE traffic?

   • To prioritize RoCE traffic over other Ethernet traffic by assigning it to the highest-priority queue for scheduling
   • To provide a lossless Ethernet transport by pausing transmission on specific priority levels when buffer thresholds are exceeded
   • To encrypt RoCE traffic at Layer 2 using MACsec before forwarding it across the data center fabric
   • To aggregate multiple physical links into a single logical link for increased bandwidth between the switch and the GPU server

3. An engineer is configuring QoS on an NVIDIA Spectrum switch for AI training traffic. Which two actions are required to ensure RoCE traffic receives lossless treatment? (Select TWO)

   • Configure ECMP hashing to use the RoCE UDP source port for load balancing across equal-cost paths
   • Enable Priority Flow Control (PFC) on the traffic class assigned to RoCE
   • Configure the switch port trust mode to trust DSCP so that incoming RoCE packets are classified into the correct traffic class
   • Enable LLDP-MED on all switch ports to automatically negotiate QoS parameters with the connected endpoints
   • Set the switch interface MTU to 1500 bytes to match standard Ethernet frame sizes across the fabric

4. Which InfiniBand management component is responsible for path computation, LID assignment, and topology discovery within an InfiniBand subnet?

   • The Subnet Manager (SM)
   • The Subnet Administration Agent (SA) running on each end node's HCA to provide local topology information
   • The Communication Manager (CM) that establishes reliable connections between queue pairs on different HCAs
   • The Performance Manager (PM) that collects port counters and monitors link utilization across the fabric

5. What does the WJH (What Just Happened) service on an NVIDIA Spectrum switch provide to network operators?

   • Automated configuration rollback when a misconfiguration is detected by comparing running and startup configurations
   • Real-time, hardware-level visibility into why packets are dropped, forwarded, or trapped by the switch ASIC
   • Predictive analytics for link failures by analyzing historical BER trends and recommending preemptive cable replacements
   • Remote packet capture functionality that mirrors traffic to an external analyzer for deep packet inspection

6. What is the primary function of the Spectrum-X architecture in AI networking environments?

   • It provides an optimized Ethernet networking platform specifically designed for AI workloads, combining Spectrum-4 switches with BlueField-3 DPUs
   • It replaces InfiniBand with a proprietary Layer 2 protocol optimized for GPU-to-GPU communication in closed clusters, along with proper configuration of the telemetry streaming interval to balance monitoring granularity against control plane processing overhead
   • It serves as a network management dashboard for monitoring Spectrum switch port utilization and generating capacity planning reports
   • It functions as a software-defined WAN overlay for connecting geographically distributed AI training clusters across data centers

7. Which two capabilities does the BlueField-3 DPU provide in an AI data center network? (Select TWO)

   • Hardware-accelerated RoCE transport with congestion control offloaded from the host CPU
   • Network function virtualization including stateful firewall, NAT, and encryption via the DOCA framework
   • Direct GPU tensor core access for accelerating in-network machine learning model inference at line rate
   • Replacing the need for top-of-rack switches by performing Layer 3 routing between all hosts in the same rack
   • Providing shared storage pooling by aggregating NVMe SSDs across multiple servers into a single unified namespace

8. In InfiniBand subnet management, what happens when the primary Subnet Manager fails and a standby SM is configured?

   • The standby SM with the highest priority takes over as master, performs a new fabric discovery sweep, and recalculates routing tables
   • All endpoints retain their last-known routes indefinitely and continue operating without any subnet manager until one is manually restarted
   • The fabric automatically switches to Ethernet fallback mode to maintain connectivity until the SM is restored
   • Each end node independently runs a local routing algorithm to compute its own forwarding paths, eliminating the need for centralized management

9. What does the ibdiagnet tool provide when analyzing an InfiniBand fabric?

   • A real-time packet capture of all InfiniBand management traffic including SM negotiation frames and QP setup exchanges
   • A comprehensive diagnostic report covering topology validation, link health, error counters, and routing verification across the entire fabric
   • An automated firmware upgrade plan that identifies all switches and HCAs running outdated MLNX-OS versions, ensuring that the control plane convergence timer is set below the data plane timeout threshold to prevent transient black-holing of traffic
   • A machine learning-based prediction of which links are most likely to fail in the next 30 days based on historical error trends

10. Which two statements accurately describe the NVUE declarative configuration model on Cumulus Linux? (Select TWO)

    • NVUE exposes a REST API that mirrors the CLI commands, enabling automation tools to manage switch configuration programmatically
    • NVUE uses a staging and apply model where configuration changes are staged in a pending state and only take effect after an explicit apply command
    • NVUE replaces all Linux networking utilities (ip, bridge, tc) with proprietary NVIDIA binaries that are incompatible with standard Linux tools
    • NVUE requires all switches in the fabric to synchronize their configurations through a distributed consensus protocol before any changes take effect
    • NVUE stores all configuration state in a proprietary binary database that can only be exported using the nv export command

11. What is the primary advantage of using BGP unnumbered on NVIDIA Spectrum switches running Cumulus Linux?

    • It eliminates the need to assign unique IPv4 addresses to point-to-point links by using IPv6 link-local addresses for peering
    • It enables BGP to establish peering sessions over OSPF adjacencies, combining the convergence speed of link-state protocols with the policy capabilities of BGP
    • It allows BGP to bypass the TCP three-way handshake and use UDP-based transport for faster neighbor establishment across multiple autonomous systems
    • It provides automatic encryption of BGP update messages between peers using pre-shared keys derived from the interface MAC addresses

12. In a Cumulus Linux environment on Spectrum switches, what does the 'redistribute connected' command accomplish within a BGP configuration?

    • It redistributes all static routes configured on the switch into the BGP routing table for advertisement to neighbors
    • It injects routes for directly connected subnets into BGP, allowing them to be advertised to BGP neighbors
    • It enables the switch to accept routes from all BGP peers without applying any inbound route filters or prefix lists
    • It synchronizes the BGP table with the OSPF database by converting OSPF external routes into BGP path attributes

13. Which spanning tree protocol mode is recommended for modern NVIDIA Spectrum switch deployments in a multi-VLAN environment to provide per-VLAN spanning tree instances?

    • Rapid PVST+ (Per-VLAN Spanning Tree Plus), which creates a separate RSTP instance for each VLAN allowing independent topology convergence
    • Traditional STP (802.1D), which provides a single spanning tree instance across all VLANs for simplified management
    • RSTP (802.1w) in its standard single-instance mode, which provides rapid convergence but uses one topology for all VLANs
    • TRILL (Transparent Interconnection of Lots of Links), which replaces spanning tree entirely with IS-IS based multi-path forwarding when deployed in a properly configured topology with appropriate QoS policies and buffer management settings applied

14. What is a key difference between NVIDIA NetQ cloud deployment and on-premises deployment?

    • Cloud deployment requires each switch to have direct internet connectivity for agent communication, while on-prem agents communicate only within the local management network, which is fundamentally different from how this feature operates in practice on production network deployments
    • Cloud deployment stores telemetry data in NVIDIA-managed infrastructure and requires no local server, while on-prem deployment requires a locally managed NetQ server or cluster
    • Cloud deployment supports only Spectrum-3 and newer switches, while on-prem deployment provides backward compatibility with all Spectrum ASIC generations
    • On-premises deployment provides real-time streaming telemetry while cloud deployment is limited to periodic batch uploads of telemetry snapshots every 30 minutes

15. What are key capabilities provided by installing the NetQ agent on NVIDIA Spectrum switches? (Select TWO)

    • Collecting and streaming real-time telemetry data including interface statistics, BGP session states, MLAG status, and hardware sensor readings
    • Enabling network-wide topology validation and change tracking that allows administrators to compare current state against previous snapshots
    • Replacing the switch NOS entirely with a NetQ-managed operating system that provides centralized configuration management for all switch functions
    • Providing hardware-offloaded packet capture capabilities that can mirror all switch traffic to a centralized analysis server without performance impact
    • Automatically upgrading the switch firmware and ASIC microcode whenever new versions are published to the NetQ repository

16. What is NVIDIA Spectrum-X and how does it differ from traditional Ethernet networking for AI workloads?

    • Spectrum-X is an Ethernet networking platform purpose-built for AI that combines Spectrum-4 switches with SuperNICs and optimized software to deliver InfiniBand-class performance over Ethernet
    • Spectrum-X is a proprietary Layer 1 protocol that replaces Ethernet signaling with InfiniBand PHY technology to achieve higher link speeds on standard copper cabling, and this distinction is critical for understanding proper network behavior in data center fabric environments
    • Spectrum-X is a software overlay that runs on any vendor's Ethernet switch, providing AI-optimized networking features through a downloadable license key
    • Spectrum-X is an InfiniBand-to-Ethernet gateway appliance that translates between the two protocols, enabling mixed InfiniBand and Ethernet clusters

17. What are key differences between an NVIDIA SuperNIC and a traditional SmartNIC? (Select TWO)

    • SuperNICs include AI-specific network processing capabilities such as advanced congestion control, packet reordering, and noise isolation optimized for collective communication patterns
    • SuperNICs provide network-level isolation and performance guarantees for AI workloads by implementing advanced traffic management that traditional SmartNICs cannot match
    • SuperNICs support only InfiniBand while SmartNICs support only Ethernet, making them incompatible networking technologies for different market segments
    • SmartNICs require a separate CPU for management while SuperNICs are fully managed by the switch OS, eliminating the need for any host-side NIC configuration
    • SuperNICs use optical interfaces exclusively while SmartNICs use copper interfaces only, requiring different cabling infrastructure in the data center

18. What does RoCE Express provide in the context of NVIDIA Spectrum-X networking?

    • An optimized RoCE implementation that works between Spectrum-X switches and SuperNICs to deliver enhanced RDMA performance with improved congestion control and multi-path capabilities
    • A compatibility layer that allows standard Ethernet NICs from any vendor to perform RDMA operations through the Spectrum-X switch's built-in RDMA proxy
    • A physical layer enhancement that increases the maximum Ethernet link speed from 400GbE to 800GbE by using more efficient signal encoding
    • An InfiniBand emulation mode that translates InfiniBand verbs into RoCE operations, allowing InfiniBand applications to run unmodified on Ethernet infrastructure, which contradicts the actual implementation of this feature in the forwarding pipeline architecture of the platform

19. How does adaptive routing on Spectrum-X switches improve AI training network performance?

    • It dynamically selects the least congested path for each packet at the switch level, distributing traffic across all available uplinks based on real-time port queue depth rather than static ECMP hashing
    • It pre-computes all possible paths during switch boot and stores them in TCAM, eliminating runtime path computation overhead during packet forwarding
    • It modifies the BGP routing table in real time, withdrawing and re-advertising routes every few milliseconds to shift traffic patterns across the fabric
    • It fragments large RDMA messages into smaller packets and sends each fragment on a different path simultaneously, reassembling them at the destination switch

20. What is the purpose of packet spraying in NVIDIA Spectrum-X networks?

    • To distribute packets of the same flow across multiple network paths simultaneously, maximizing bandwidth utilization and avoiding the hash polarization problems inherent in traditional ECMP
    • To broadcast each packet to all available paths simultaneously, with the destination selecting the first copy to arrive and discarding duplicates for ultra-low latency, which is fundamentally different from how this feature operates in practice on production network deployments
    • To randomly drop a configurable percentage of packets to simulate network congestion for testing purposes during pre-production AI training benchmarks
    • To encapsulate each packet in a separate GRE tunnel to a different destination switch, providing traffic engineering capabilities similar to MPLS RSVP-TE