NVIDIA Certified Professional - AI Networking practice questions

Question 1

An organization is cabling a new QM9700 switch and the data-center team needs the per-port speed range and the connector cage type before deployment. Per the QM97XX user manual, which option is correct?

Accepted Answer

Each port runs at 40 to 400 Gb/s and the chassis exposes a connector cage of 32 OSFP ports.

Answer

Each port runs at 10 to 100 Gb/s and the chassis exposes 64 QSFP-DD cages across its front panel.

Answer

Each port runs at a fixed 200 Gb/s and the chassis exposes 40 SFP-DD cages across its front panel.

Answer

Each port runs at 25 to 400 Gb/s and the chassis exposes 36 OSFP-RJ45 hybrid cages on its front panel.

Question 2

A security engineer is tasked with hardening a Cumulus Linux switch and is building an ACL policy file for cl-acltool. Per the Cumulus Linux 5.16 Access Control List Configuration documentation, which TWO of the following are valid rule categories a rule can be assigned to in the policy file? (Select two.)

Accepted Answer

[iptables] for IPv4 packet-filter rules processed by cl-acltool

Accepted Answer

[ebtables] for layer-2 Ethernet bridge filtering rules in the file

Answer

[arptables] as the dedicated ARP-filtering category cl-acltool reads

Answer

[nftables] as the default category cl-acltool generates rules into

Question 3

A cloud administrator is tasked with capturing per-process NCCL debug logs from a multi-node training job into separate files so the output from different ranks does not clobber each other. According to the NCCL documentation, how should NCCL_DEBUG_FILE be set?

Accepted Answer

Set NCCL_DEBUG_FILE to a filename.%h.%p pattern so %h and %p expand to host and PID, keeping each rank's log unique.

Answer

Point NCCL_DEBUG_FILE at one shared path; NCCL appends each rank's lines so the merged log stays ordered and intact.

Answer

Set NCCL_DEBUG_FILE to a ~ home-relative path; NCCL expands the tilde per host so each rank writes its own log file.

Answer

Leave NCCL_DEBUG_FILE unset and rely on NCCL_DEBUG_SUBSYS alone to route each rank's output to its own per-host file.

Question 4

A production team is planning a fabric-wide Cumulus Linux upgrade through NetQ lifecycle management and must decide the order in which switches are upgraded based on their assigned roles. Per the NetQ Switch Management documentation, which option is correct?

Accepted Answer

The upgrade begins with superspine switches, then spine, then leaf and exit switches, and finally switches with no role.

Answer

The upgrade begins with leaf switches first, so access-layer changes are validated before any spine switch is ever touched.

Answer

NetQ upgrades all four switch roles simultaneously and rolls the whole fabric back if any single switch fails its upgrade.

Answer

The upgrade order is alphabetical by hostname, since role assignments only affect telemetry collection and not upgrades.

Question 5

An architect is designing a DGX SuperPOD and must explain to leadership the architectural ingredients beyond the DGX systems themselves. Per the DGX SuperPOD architecture documentation, which option correctly lists the ingredients?

Accepted Answer

The DGX SuperPOD architecture is a combination of DGX systems, InfiniBand and Ethernet networking, management nodes, and storage

Answer

The DGX SuperPOD architecture is purely DGX systems with no management or storage tier

Answer

The DGX SuperPOD architecture is built around a single Wi-Fi mesh and no InfiniBand or Ethernet

Answer

The DGX SuperPOD architecture uses MPLS WAN circuits as the primary intra-pod fabric

Question 6

A developer is writing the Pod spec for a training job and wants to reserve two NVIDIA GPUs that the device plugin advertised as nvidia.com/gpu. A teammate warns that GPUs follow stricter rules than cpu and memory. According to the Kubernetes GPU scheduling guidance, how should the GPU resource be expressed?

Accepted Answer

Put nvidia.com/gpu under limits; if requests is omitted the limit is used as the request, and both values must be equal.

Answer

Put nvidia.com/gpu only under requests, since for custom devices Kubernetes simply ignores whatever value you place in the limits field.

Answer

Put nvidia.com/gpu under requests and a smaller value under limits so the scheduler is free to overcommit the GPUs across the pods.

Answer

Put nvidia.com/gpu into a node selector label, because GPUs are not permitted to appear in a container's resources block at all.

Question 7

An organization is scaling an all-to-all workload with sparse, bursty traffic and wants to cut the number of queue pairs each node keeps open. Per the MLNX_OFED Advanced Transport documentation, which option is correct?

Accepted Answer

Dynamically Connected transport lets Reliable QPs connect and disconnect on demand, cutting the QP count.

Answer

Extended Reliable Connected transport pins one static QP per peer, which suits sparse traffic best of all.

Answer

Unreliable Datagram transport removes the QP concept fully, so no connections are set up between any nodes.

Answer

Shared Receive Queues pool all inbound buffers, removing the need for per-connection queue pairs entirely.

Question 8

A company is evaluating Spectrum-X for a noisy-neighbor problem in which one training job's incast bursts are stealing throughput from a colocated inference workload. Per the Spectrum-X congestion-control documentation, what mechanism provides per-workload performance isolation?

Accepted Answer

Telemetry-based congestion control combines high-frequency telemetry probes with flow metering for performance isolation across workloads

Answer

Spectrum-X uses a single 8-MB shared buffer per ASIC and disables PFC so any one tenant naturally backpressures every other tenant in the pod

Answer

Spectrum-X requires every job to run inside its own physical leaf pod so no congestion signal ever crosses workloads on the shared fabric

Answer

Spectrum-X falls back to MPLS traffic-engineered LSPs to reserve dedicated bandwidth per workload before any AllReduce phase can start

Question 9

A sysadmin is tasked with gathering diagnostic data from a Cumulus Linux switch after an application produced a core dump, and wants to know when the support bundle is produced automatically. According to the Cumulus Linux documentation, which statement about cl-support is correct?

Accepted Answer

cl-support builds a compressed archive for troubleshooting and is generated automatically when an application core dumps.

Answer

cl-support only exports a live sFlow sample stream to a collector and never writes any local archive you could attach.

Answer

cl-support runs only when invoked by hand, so the switch never builds the bundle when a monitored service first fails.

Answer

cl-support writes an uncompressed log directory and is built only when an InfiniBand port drops to the Init link state.

Question 10

A platform team is rolling out Cumulus Linux on a new switch fleet, and an engineer asks how the NVUE CLI command structure differs from a traditional modal network CLI. Per the Cumulus Linux NVUE CLI documentation, which option is correct?

Accepted Answer

The NVUE CLI uses a flat structure, letting you run all NVUE commands from the primary prompt instead of entering a configuration mode.

Answer

The NVUE CLI uses a modal structure, so you must first enter a dedicated configuration mode before the switch will accept any nv command.

Answer

The NVUE CLI accepts nv commands only after you launch the vtysh routing shell and enter its privileged enable mode first.

Answer

The NVUE CLI needs a separate SSH login per command because every nv command opens and then closes its own session mode.

Question 11

An organization is scaling an InfiniBand deployment past the limits of a single subnet and must connect several fabric islands that use different topologies. Which statement correctly describes what an InfiniBand router provides in this scenario?

Accepted Answer

An IB router connects two or more subnets and lets the fabric scale beyond 40,000 end-ports, forwarding cross-subnet traffic via Global Routing Header lookup.

Answer

An IB router merges the separate subnets into one large subnet under a single subnet manager, which is what lets the fabric exceed the 40,000 end-port ceiling.

Answer

An IB router bridges InfiniBand to Ethernet, so the fabrics must first be converted to RoCE before any cross-subnet traffic can be forwarded between the islands.

Answer

An IB router only links subnets that already share an identical topology, so Fat Tree and DragonFly+ islands cannot be interconnected without rebuilding one of them.

Question 12

An engineer is tuning a latency-sensitive inference deployment and sets the kubelet's Topology Manager policy to single-numa-node so each pod's GPU, NIC, and CPU come from the same NUMA node. During the rollout several pods stay Pending and never schedule on otherwise-idle GPU nodes. Given how the single-numa-node policy works, what is the most likely explanation?

Accepted Answer

Under the single-numa-node policy, Topology Manager will reject the pod from a node whenever a single NUMA node affinity cannot be achieved.

Answer

The kubelet quietly downgrades the affinity to best-effort, so a pod only stays Pending once the cluster has run completely out of free GPUs.

Answer

Topology Manager spreads each GPU allocation across all NUMA nodes by default, so the policy fails only when the device plugin never registered.

Answer

Topology Manager merges resource hints from every NUMA node, so a Pending pod must instead stem from missing CPU requests in the pod spec.

Question 13

An MLOps team is being asked to quantify how much fabric bandwidth a training job loses specifically to congestion, and the engineer must map UFM's calculated counters to that need. Per the UFM port-counters documentation, which counter captures it?

Accepted Answer

Normalized Congested Bandwidth: bandwidth suppressed by congestion, from (XmitWait incremental / Time) * Link Capacity.

Answer

Normalized XmitData, which reports the effective port bandwidth utilization as a percentage of the port's link capacity.

Answer

XmitWait, the raw count of ticks the port spent unable to transmit for any reason during the sampling interval.

Answer

SymbolErrorCounter, the count of physical-layer symbol errors observed on the congested link during the run.

Question 14

A data center architect is designing an EVPN fabric on Cumulus Linux and must decide between distributed symmetric and distributed asymmetric inter-subnet routing. Per the Cumulus Linux inter-subnet routing documentation, which statement correctly describes distributed symmetric routing?

Accepted Answer

Every VTEP participates in routing and performs routing at both the ingress and the egress VTEP

Answer

Specific VTEPs act as designated layer 3 gateways while the remaining VTEPs only perform bridging

Answer

Every VTEP routes, but all routing happens at the ingress VTEP and the egress VTEP only bridges

Answer

Only spine switches route between subnets while leaf VTEPs bridge VLANs across VXLAN tunnels

Question 15

An organization is evaluating NetQ as the telemetry backbone for a Spectrum-X fabric that also has DPUs and NVLink links to observe. Per the NetQ Components documentation, which option is correct?

Accepted Answer

NetQ collects telemetry from Ethernet switches, DPUs, hosts, and NVLink fabrics, normalizes it, and streams it into a central analytics layer

Answer

NetQ collects telemetry only from Cumulus Linux switches; DPUs, hosts, and NVLink are out of scope

Answer

NetQ collects telemetry only from NVLink-attached GPUs; switches and DPUs are unsupported

Answer

NetQ collects raw syslog into a flat file on the controller node with no normalization, no correlation, and no streaming pipeline whatsoever

NVIDIA Certified Professional - AI Networking

Sample questions

Sources

Similar exams