AI Wisdom

NVIDIA professional level

NVIDIA Certified Professional: AI Networking

Code
NCP-AIN
Length
120 minutes
Price
$400. NVIDIA says labs are added on November 1, when it rises to $500.
Valid for
2 years, renewed by retaking the exam
Study time, our estimate
90 to 130 hours

A specialist exam for network engineers moving into AI fabrics. Ethernet and InfiniBand get equal weight, and tools another fifth. If you already hold strong routing and switching skills, this is a sensible way to prove you can carry them over.

Take it if

You run the network under GPU clusters.

Skip it if

Networking is not your job. Few AI engineers need this depth.

Before you start

Check yourself against this list first. If something is missing, close that gap before you pay for an exam.

  • Two to three years of data center workNVIDIA recommends this.
  • Strong Ethernet and BGPRouting and switching are assumed.
  • InfiniBand or RoCE exposureBoth are expected.

Helpful first: NCA-AIIO

All exam details
Format
70 to 75 questions, pass or fail
Length
120 minutes
Price
$400. NVIDIA says labs are added on November 1, when it rises to $500.
Valid for
2 years, renewed by retaking the exam
Delivery
Proctored, English
Experience
2 to 3 years of hands-on data center networking, ideally with NVIDIA networking hardware
If you fail
Wait 14 days

What the exam covers

  • NVIDIA Spectrum Networking30%
  • NVIDIA InfiniBand Networking30%
  • Troubleshooting Tools20%
  • Automation and Configuration10%
  • AI Data Center Design and Optimization5%
  • Kubernetes Integration5%

What to focus on

  • Spectrum Ethernet and Cumulus Linux configuration.
  • InfiniBand: subnet managers, fabric design and administration.
  • RoCE and RDMA, and lossless Ethernet.
  • Fabric diagnostic tools.
  • Automation, plus Kubernetes network integration.
Everything on one mapEvery topic and key idea in this certification.
mindmap
  root((NCP-AIN))
    Spectrum Ethernet
      RoCE
      PFC and ECN
      Spectrum-X
    InfiniBand
      Subnet manager
      UFM
      Topology
    Troubleshooting tools
      ibdiagnet
      Counters
      Method
    Automation and design
      Automation
      Kubernetes
      Design
Why AI traffic needs a lossless fabricA simplified view of how RoCE congestion is handled.
sequenceDiagram
  participant S as Sender
  participant W as Switch
  participant R as Receiver
  S->>W: RDMA traffic
  W->>R: Marks packets with ECN as queues build
  R-->>S: Congestion notification
  S->>S: Slow down
  W-->>S: PFC pause if the buffer is nearly full

Lesson 1 of 4, 30% of the exam

Spectrum Ethernet

30% of the exam.

RoCE
RDMA over Ethernet needs lossless behavior.
PFC and ECN
Prevent drops and signal congestion.
Spectrum-X
NVIDIA's Ethernet platform for AI.
Lossless Ethernet for RoCETwo mechanisms working together.
flowchart LR
  A["Queues start to build"] --> B["ECN marks packets"]
  B --> C["Sender slows down"]
  A --> D["Buffer nearly full"]
  D --> E["PFC pauses the sender"]

Common trap

Treating AI traffic like general data center traffic.

Try it yourself

Sketch a RoCE configuration for a leaf switch.

Official docs: Spectrum-X networking

Lesson 2 of 4, 30% of the exam

InfiniBand

Equal weight at 30%.

Subnet manager
Assigns addresses and programs routes.
UFM
Central fabric management.
Topology
Design for the traffic.
What the subnet manager doesNo subnet manager means no traffic.
sequenceDiagram
  participant S as Subnet manager
  participant P as Switches and adapters
  S->>P: Discover the fabric
  S->>P: Assign addresses
  S->>P: Program routes
  P-->>S: Ports go active

Common trap

Forgetting the subnet manager when ports stay down.

Try it yourself

Take InfiniBand Essentials on NVIDIA Academy.

Official docs: NVIDIA training

Lesson 3 of 4, 20% of the exam

Troubleshooting tools

A fifth of the exam.

ibdiagnet
Scans the fabric for errors.
Counters
Read link errors.
Method
Start physical, then move up.
Working up the stackMost fabric faults start low.
flowchart TD
  A["Physical: cable and optic"] --> B["Link: port state and counters"]
  B --> C["Fabric: ibdiagnet scan"]
  C --> D["Application"]

Common trap

Using ping alone on a fabric problem.

Try it yourself

Read sample diagnostic output.

Official docs: NVIDIA training

Lesson 4 of 4, 10% + 5% + 5% of the exam

Automation and design

Twenty percent, and it scales everything else.

Automation
Versioned config prevents drift.
Kubernetes
The exam includes network integration.
Design
Plan for AI traffic.
Config as codeOne source of truth for every switch.
flowchart LR
  A["Template in Git"] --> B["Review"]
  B --> C["Automation"]
  C --> D["Leaf switches"]
  C --> E["Spine switches"]

Common trap

Configuring 200 switches by hand.

Try it yourself

Write one switch config as a template.

Official docs: Spectrum-X networking

Study plan

  1. Weeks 1 and 2: InfiniBand

    Take InfiniBand Essentials and InfiniBand Network Administration on NVIDIA Academy.

  2. Weeks 3 and 4: Ethernet

    Take Cumulus Linux Essentials and configure a virtual fabric.

  3. Week 5: troubleshoot

    Learn the diagnostic tools until you can read their output quickly.

  4. Exam week

    Check the exam page for the lab format, then take the mock below.

Exam day and after

  • NVIDIA certifications last two years, and you renew by retaking the exam.
  • Exams are proctored and in English. NVIDIA's FAQ covers both remote and in-person testing, and remote exams allow no breaks.
  • Check the exam page before you plan around a date. NVIDIA has been adding hands-on labs to professional exams.
  • If you book after the lab change, plan practice time on a real or virtual fabric, because reading will not be enough.
Test yourself8 practice questions, scored by exam topic.
Start the mock exam

Keep going

Free on AI Wisdom

Exam facts checked on October 31, 2026 against NVIDIA certification page. Lessons are our own explanations, so use the official docs as the final word, and confirm price and dates before booking.