Enroll

thinQmaster

Find your exam

Five questions free, before you decide.

Step 1 · choose a vendor

In one of our camps? Your practice exams come with the tuition — they are on your study page already.

thinQmaster — free preview

AI Networking (NCP-AIN). Set it up and run it.

This is not a sample sheet — it is the application, running. Pick your mode, your length, your domain, the same way you would inside the full title. Study mode marks each answer and tells you why the right one is right and why each wrong one is wrong; Simulation puts a clock on it and chooses each question from whatever you are weakest in. No account, no e-mail address, no card.

Free · nothing to sign up for

thinQmaster — AI Networking (NCP-AIN)

Built like the exam, not like a flashcard deck. The same question shapes you meet on test day, sat against the published blueprint, with a timed paper at full length.

508 questions in the full guide

Question types

475 multiple choice · 33 choose two

Domains, and how many questions each

  • Domain 1 — Design and Architecture of AI Networks25
  • Domain 2 — NVIDIA Spectrum-X Ethernet Networking150
  • Domain 3 — NVIDIA InfiniBand Networking150
  • Domain 4 — Kubernetes Integration25
  • Domain 5 — Troubleshooting and Validation108
  • Domain 6 — Network Automation50

A real exam experience

  • Timed simulation — 72 questions in 120 minutes, weighted to the blueprint, no feedback until the end.
  • 6 domain drills — work one weak area at a time.
  • Exhibits — real command output and topology diagrams to read, the way the exam asks them.
  • It adapts — miss a domain and the next questions lean that way; master one and it eases off.
  • Every answer explained — why the right one is right, and why each wrong one is wrong.

And what comes with it

  • Twelve months of access — and of updates. Blueprints change and questions get added; while your year is running you get every one of them at no extra cost.
  • A certified engineer on e-mail. Stuck on why an answer is what it is? thinqmaster@thinqtanklearning.com reaches somebody who holds the certification and works in the field — not a ticket queue.
  • The discussion groups. Post a question to other people sitting the same exam and get it talked through.
  • Not a subscription. You sit the exam once; you pay once, $99. Nothing renews and nothing has to be cancelled.
  • In one of our camps? It is already included in your tuition — you will find it in your classroom.

Below is the engine itself, running on real AI Networking (NCP-AIN) questions. Free, no account, no card.

thinQmaster engine · preview session

The five questions

1. An architect is separating the networks for a new GPU cluster. Which traffic belongs on the compute (east-west) fabric?

  • A. User SSH sessions to the login nodes
  • B. Dataset reads from the parallel file system
  • C. GPU-to-GPU collective traffic between servers during training
  • D. BMC and switch management traffic

Correct answer: C. The compute fabric carries inter-server GPU traffic such as all-reduce.

Why the others are wrong. Incorrect choices: A) User SSH sessions to the login nodes; B) Dataset reads from the parallel file system; D) BMC and switch management traffic. Dataset reads use the storage/converged fabric; BMC traffic is out-of-band management; SSH is in-band management.

2. Which TWO are applied by nv set qos roce in lossless mode? (Choose two.)

  • A. Disabling ECN
  • B. PFC on switch-priority 3
  • C. Trust of PCP and DSCP markings
  • D. Moving RoCE to strict priority 7
  • E. Dropping CNPs

Correct answer: B. The profile enables PFC on priority 3 and trusts PCP/DSCP. It enables (not disables) ECN, does not use priority 7 for RoCE, and gives CNPs priority 6.

Why the others are wrong. Incorrect choices: A) Disabling ECN; D) Moving RoCE to strict priority 7; E) Dropping CNPs. Each describes a different tool, setting or behaviour from the one the question asks about.

3. In an HGX/DGX H100 reference design, what is the ratio of compute NICs to GPUs?

  • A. 1 NIC per server
  • B. 1 NIC per 4 GPUs
  • C. 1 NIC per 2 GPUs
  • D. 1 NIC per GPU

Correct answer: D. Each GPU has its own compute NIC (8 GPUs, 8 ConnectX-7) so every GPU gets full network bandwidth. The other ratios would oversubscribe GPUs on the network.

Why the others are wrong. Incorrect choices: A) 1 NIC per server; B) 1 NIC per 4 GPUs; C) 1 NIC per 2 GPUs. Each describes a different tool, setting or behaviour from the one the question asks about.

4. Which two elements together make up the Spectrum-X platform end to end?

  • A. Any Ethernet switch and any RDMA NIC
  • B. Spectrum-4 switches and BlueField-3 or ConnectX-8 SuperNICs
  • C. Quantum-2 switches and ConnectX-6 NICs
  • D. UFM and NetQ

Correct answer: B. Spectrum-X pairs Spectrum-4 switches with SuperNICs and treats them as one system. Quantum-2 is InfiniBand, arbitrary hardware lacks the joint features, and UFM/NetQ are management tools.

Why the others are wrong. Incorrect choices: A) Any Ethernet switch and any RDMA NIC; C) Quantum-2 switches and ConnectX-6 NICs; D) UFM and NetQ. Each describes a different tool, setting or behaviour from the one the question asks about.

5. What prevents congestion drops on an InfiniBand link by design?

  • A. TCP retransmission
  • B. Credit-based link-level flow control: a sender transmits only when the receiver has advertised buffer credits
  • C. Adaptive routing
  • D. ECN marking

Correct answer: B. IB is lossless through credits. ECN and AR exist but do not make links lossless; TCP is not used.

Why the others are wrong. Incorrect choices: A) TCP retransmission; C) Adaptive routing; D) ECN marking. Each describes a different tool, setting or behaviour from the one the question asks about.

6. What does the NVIDIA Network Operator do in a Kubernetes cluster?

  • A. Schedules GPU jobs
  • B. Replaces kube-proxy
  • C. Runs UFM
  • D. Deploys and manages the networking components needed for RDMA and InfiniBand workloads (drivers, device plugins, CNIs, IPAM)

Correct answer: D. It automates the network stack for accelerated workloads. Scheduling is the scheduler/GPU Operator; it does not replace kube-proxy or run UFM.

Why the others are wrong. Incorrect choices: A) Schedules GPU jobs; B) Replaces kube-proxy; C) Runs UFM. Each describes a different tool, setting or behaviour from the one the question asks about.

7. What does cl-resource-query report on a Spectrum switch running Cumulus Linux?

  • A. PFC pause counters
  • B. BGP neighbour state
  • C. Hardware forwarding resource usage (routes, hosts, ACL entries, ECMP groups) against ASIC limits
  • D. CPU temperature only

Correct answer: C. Cl-resource-query shows ASIC table utilization. BGP and PFC are shown by other commands.

Why the others are wrong. Incorrect choices: A) PFC pause counters; B) BGP neighbour state; D) CPU temperature only. Each describes a different tool, setting or behaviour from the one the question asks about.

8. What role does the BlueField-3 DPU typically play in each server of an AI factory?

  • A. It terminates the storage/north-south network and offloads infrastructure services such as tenant isolation and security
  • B. It is the compute-fabric NIC for GPU 0
  • C. It replaces NVSwitch for GPU-to-GPU traffic
  • D. It runs the InfiniBand Subnet Manager

Correct answer: A. The DPU fronts the storage/converged network and runs infrastructure services.

Why the others are wrong. Incorrect choices: B) It is the compute-fabric NIC for GPU 0; C) It replaces NVSwitch for GPU-to-GPU traffic; D) It runs the InfiniBand Subnet Manager. NVSwitch handles intra-server GPU traffic; the SM runs on switches, hosts or UFM; compute NICs are ConnectX/SuperNICs.