Design the AI fabric
Describe an AI factory’s network, choose between Ethernet and InfiniBand, and lay out rail-optimized topologies.
Partners and press
Overview
GPUs are only as fast as the fabric between them. This exam tests the fabric.
NCP-AIN is the Professional exam on AI networking for engineers working with NVIDIA hardware, with two to three years advised. Spectrum-X Ethernet and InfiniBand carry 30% each: configuration, quality of service, telemetry and multi-tenant isolation. NVIDIA says it includes a hands-on lab, and the rest covers topologies, RDMA in Kubernetes, congestion tracing and automation.
12Modules
70–75Questions
120Minutes
8Roles
What it tests
Learn all the way down.
Learn all the way down.
Perform when it matters.
Run two fabrics
Configure Spectrum-X for RoCE and InfiniBand with partition keys, QoS, adaptive routing and tenant isolation.
Trace congestion
Use NetQ, UFM, What Just Happened and the InfiniBand diagnostic tools to find where packets slow or drop.
Automate the switches
Manage Spectrum-X configuration with NVUE templates and write Ansible playbooks for VLANs and RoCE.
Course outline
From a rail-optimized design
From a rail-optimized design
to a fabric you can automate.
We design this course ourselves. NVIDIA runs the exam and issues the credential, and LUNARTECH is independent of NVIDIA.
Training and inference put unusual loads on a network. You learn how collective operations create bursts, why east–west bandwidth dominates, and how compute, storage and management fabrics divide the work.
- Collectives and their traffic patterns
- East–west and north–south traffic
- Compute, storage and management fabrics
- Latency, bandwidth and tail latency
- GPU-to-GPU paths through the network
- What congestion costs a training job
You build A traffic model for one training job across a cluster.
Both fabrics can carry AI, and each suits a different situation. You compare them on performance, operations, ecosystem and cost, and reach a recommendation you can defend to a team.
- RDMA over Converged Ethernet in context
- InfiniBand’s native features
- Lossless behavior on each fabric
- Operations and tooling compared
- Multi-tenancy and cloud considerations
- Mixed fabrics in one AI factory
You build A written comparison and recommendation for one deployment.
Topology decides how many hops a gradient takes. You design rail-optimized and fat-tree layouts, calculate oversubscription, plan for scale and redundancy, and cable the result so failures stay contained.
- Fat-tree and leaf–spine layouts
- Rail-optimized design for GPU servers
- Oversubscription and blocking
- Scaling out in pods and tiers
- Redundancy and failure domains
- Cabling plans and port maps
You build A topology design with a port map for a small AI cluster.
RoCE needs a network that does not drop packets. You configure traffic classes, priority flow control and ECN, see how congestion control reacts, and learn the misconfigurations that stall a fabric.
- How RoCE moves data
- Traffic classes and DSCP mapping
- Priority flow control and its dangers
- ECN marking and congestion control
- Buffer behavior and headroom
- Common RoCE misconfigurations
You build A RoCE configuration for a switch pair, tested under load.
Spectrum-X is NVIDIA’s Ethernet platform for AI. You configure Spectrum switches for routing, adaptive routing and QoS, and add BGP and EVPN to carry several tenants over the same fabric.
- Spectrum-X platform and its parts
- Cumulus Linux and the NVUE model
- BGP routing for the underlay
- Adaptive routing and congestion handling
- BGP-EVPN and VXLAN overlays
- Microsegmentation between tenants
- Securing switch access
You build A multi-tenant Spectrum-X fabric with isolated tenants and tuned QoS.
A fabric you cannot see will surprise you. You collect telemetry with NetQ, use What Just Happened to see dropped packets and the reasons behind them, and turn the data into alerts and baselines.
- Telemetry sources on Spectrum switches
- NetQ for validation and history
- What Just Happened for packet drops
- Counters, buffers and queue depth
- Baselines and alert thresholds
- Streaming telemetry to dashboards
You build A monitoring setup that flags drops and congestion on a Spectrum fabric.
InfiniBand is managed centrally. You start a subnet manager, assign partition keys for isolation, choose routing and adaptive routing, and use in-network reduction where the fabric supports it.
- InfiniBand addressing: LIDs and GIDs
- The subnet manager and its failover
- Partition keys and tenant isolation
- Routing engines and adaptive routing
- Quality of service and virtual lanes
- In-network reduction with SHARP
You build An InfiniBand fabric with a subnet manager, partitions and verified routing.
UFM is the control point for large InfiniBand fabrics. You discover topology, read port and link health, find congestion hot spots, and use alerts and health tests to stop small faults spreading.
- UFM architecture and deployment
- Discovering and visualizing the fabric
- Telemetry and congestion maps
- Events, alarms and health tests
- Managing partitions from UFM
- Upgrading and backing up UFM
You build A UFM deployment that maps a fabric and raises alerts on link errors.
Containers need the same fast paths as bare metal. You deploy the Network Operator, attach secondary RDMA networks to pods, and test that GPU workloads in Kubernetes really use the fabric.
- What the Network Operator installs
- Secondary networks and multiple interfaces per pod
- SR-IOV and shared RDMA devices
- Running RDMA and NCCL tests in pods
- Network policy and isolation
- Upgrading operator components
You build A Kubernetes cluster where pods use RDMA over a secondary network, verified with tests.
When training slows, the network is the first suspect and rarely the whole story. You use diagnostic tools to measure paths, read counters and logs, and isolate the links, optics and settings that cause trouble.
- ibdiagnet fabric checks
- Bandwidth and latency tests with perftest tools
- Reading link errors and flapping ports
- What Just Happened and drop reasons
- UFM health and event review
- From an NCCL slowdown to a bad link
- Writing up a root cause
You build A diagnostic runbook, tested on faults you inject into a fabric.
Hand-configured switches drift. You describe configuration as NVUE templates, apply it with Ansible, provision new switches without touching them, and check that every device still matches the design.
- The NVUE object model and CLI
- Writing and applying NVUE templates
- Ansible inventories and playbooks
- Automating VLANs and RoCE settings
- Zero-touch provisioning
- Validating configuration against intent
- Keeping configurations in version control
You build Playbooks that configure VLANs and RoCE across a set of switches from one source.
Design a network for a stated GPU deployment, configure it for RoCE and InfiniBand in a lab, monitor it, break it and fix it, then automate the result so it can be rebuilt on demand.
- Reading a requirement and sizing the fabric
- Configuring the fabric end to end
- Isolating tenants and verifying QoS
- Tracing an injected congestion fault
- Automating the full build
- Presenting the design
You build A designed, configured, monitored and automated AI fabric with test evidence.
Issued by
Certifying body · Independent of LUNARTECH
NVIDIA
NVIDIA designs the GPUs and much of the software that AI runs on. It runs this certification program itself: the exams are online and proctored remotely, and passing earns a digital badge. LUNARTECH is independent of NVIDIA.
How you get there
Top-tier roles go to people who can do the work.
Top-tier roles go to people who can do the work.
Here’s how you get there.
Know where you start.
NVIDIA recommends two to three years of hands-on data center networking experience, ideally with NVIDIA networking hardware.
Work through the course.
Twelve modules, designed by LUNARTECH, at your own pace.
Sit it with NVIDIA.
NVIDIA runs the exam and issues the credential, so you register and book directly with them.
Support
Support at every stage.
Support at every stage.
You won’t be working through this alone.
Before you start
Try it first.
A 3-day free trial opens a selection of the content. Your card is only charged once it ends.
While you learn
Mentors within reach.
Stuck on a concept? Reach out to a mentor whenever you need help.
While you learn
One-on-one sessions.
On the yearly plan, one-on-one sessions give you dedicated support.
Any time
A team behind you.
Write to info@lunartech.ai at any point, from your trial onwards. If your employer is funding you, we’ll set it up with them directly.
All the way through
A schedule that fits around your life.
Your own paceFrom anywhereAt any hour
After the last lesson
Stay in the network.
Your alumni network stays with you: ongoing support, opportunities and industry insight.
Career support
Career support, past the last lesson. Coaching, a sharper resume, real interview preparation and the people to know.
Career direction
Tailored coaching helps you decide where to aim and how to get there.
Strong portfolio
Hands-on projects become work you can show, not just claim.
Resume and profile
Personalized refinement, so what you send says what you can actually do.
Interview preparation
Preparation for the questions that decide it, so you walk in ready.
Network and search
Connections with industry leaders, and an alumni network that stays with you.
Reviews
Real stories from the community.
LUNARTECH Academy has been the most empowering learning experience of my life, giving me a clear, structured path from the absolute basics to building real-world AI systems with confidence.
Your page is fantastic. As a bachelor’s student in Data Science now accepted for a master’s at UConn, your work has been a real source of motivation.
The programme’s clear, bite-sized content and practical projects dramatically improved my speed from fuzzy idea to MVP and my ability to communicate AI value. I followed the system, and it works.
I really appreciate your teaching style and the way you break down complex topics. It has been incredibly helpful.
After years working mainly with R for econometrics, I found your Python machine learning course a refreshing and engaging way to learn Python.
As a Data Science graduate, I want to thank you for the effort you put into making videos to support growing data scientists.
I’ve been following your work since taking your machine learning course, and I’m impressed by your practical approach to applying AI.
Why choose us
Built around how you actually learn.
One standard behind every course, from the first lesson to long after the last.
Hands-on practice
Apply every idea to real problems, so it becomes something you have used, not only seen.
Industry-aligned curriculum
We keep courses current with the tools, techniques and skills the field asks for.
Flexible learning
Study at your own pace, from anywhere, at any hour.
Certificate of completion
Finish the course and earn a certificate that shows what you have learned.
Focus on your success
Personalized learning plans and support structures carry you through the hard parts.
Lifetime alumni network
Stay connected after you finish, with ongoing support, opportunities and industry insight.
Your certificate
It doesn’t raise its voice. Neither will you.
Finish a LUNARTECH course and its certificate is issued in your name, under our signature. NVIDIA’s credential comes from NVIDIA, when you pass its exam.
Certificate of completion
Your Name
NCP-AIN course
LUNARTECH RISE Academy
Pricing
This course comes with the membership.
This course comes with the membership.
So does everything else.
One membership opens every course, certification, workshop and masterclass in the Academy. Every plan starts with a 3-day free trial. You register and book the exam directly with NVIDIA.
3 days free, then $29.97 a month. Your card is only charged once the trial ends.
Included in every plan
- Every course in the Academy
- Every certification course
- Workshops and masterclasses
- A certificate of completion
- A 3-day free trial
- Learning at your own pace
- The lifetime alumni network
- Upgrades at any time
In practice
Eight roles.
Eight roles.
One credential across them all.
No single job owns it. Here’s where each role leans hardest, and where its work ends up.
- 01Design
- 02Spectrum-X
- 03InfiniBand
- 04Kubernetes
- 05Troubleshooting
- 06Automation
-
Packet loss slows every GPU in the job. Telemetry shows where the loss starts.
Spectrum-X, InfiniBand, Troubleshooting
AI data centers, Enterprise IT
-
Topology sets how well a cluster scales. Get the design right before the cabling.
Design, Kubernetes
AI data centers, Research computing
-
Hundreds of switches can’t be configured by hand. Templates and playbooks keep them identical.
Automation, Spectrum-X
Cloud providers, Colocation facilities
-
Topology sets how many GPUs can talk at full speed. Ethernet or InfiniBand is a design decision.
Design, Spectrum-X, InfiniBand
AI data centers, Cloud providers
-
Congestion shows up as slow training, not as a down alert. Telemetry shows where the traffic piles up.
Troubleshooting, Kubernetes
Cloud providers, Enterprise IT
-
Multi-tenant clusters need isolation at the network layer. QoS and tenant boundaries keep neighbors apart.
Spectrum-X, InfiniBand
Colocation facilities, Cloud providers
-
Switch configs belong in templates, not terminals. NVUE and Ansible make changes repeatable.
Automation, Kubernetes
Cloud providers, Enterprise IT
-
RDMA in Kubernetes gives pods direct network access. Getting that setup right is a platform job.
Kubernetes
Cloud providers, Enterprise IT
- AI data centers
- Cloud providers
- Research computing
- Colocation facilities
- Enterprise IT
Network Engineer
Packet loss slows every GPU in the job. Telemetry shows where the loss starts.
Over 10,000 students trust LUNARTECH
LUNARTECH gives you the skills
to learn, build, create and lead.
Engineers
Understand systems all the way down. Build the ones others rely on.
Scientists
Understand what the data says. Build models that hold up.
Researchers
Understand the papers. Build on them, and take the work further.
The field
Anyone can use AI.
Anyone can use AI.
Few can engineer it. This is where that starts.
+340%growth in open roles requiring AI skills
Roles and pay
~35,000 open roles right now
Open roles: Built In, Greenhouse, LinkedIn.
Typical salaries: Glassdoor, Levels.fyi, Payscale, Indeed, Ravio, JobRise, U.S. BLS. Military: 2026 U.S. basic pay, drone officers with aviation incentive pay.
Updated November 2026.
FAQ
Frequently asked questions.
Need help? These answers cover most of what people ask before they start.
No. NVIDIA issues it. LUNARTECH is independent of NVIDIA, and its names and marks belong to it. When you finish a course of ours, you earn a LUNARTECH certificate of completion for that course.
No. We design this course ourselves, around what the exam covers. NVIDIA has no part in it, and you still register and book the exam directly with NVIDIA.
70–75 questions, 120 minutes, in English.
NVIDIA recommends two to three years of hands-on data center networking experience, ideally with NVIDIA networking hardware.
2 years. Retake the exam.
We last reviewed it on 5 November 2026, from NVIDIA’s own pages. Details change, so check the official page before you book.
Our 3-day free trial gives you access to a selection of content in the bootcamps, courses and materials, so you can experience the value of our content first-hand. This includes comprehensive learning materials and interactive sessions. You won’t be charged during this period, though we require your credit card details at the start to ensure a seamless transition should you choose to continue.
Yes, entering your credit card details is a necessary step to begin your free trial. The process is entirely secure, and we ask for it to provide a seamless transition to a paid subscription if you decide to continue beyond the trial. You will only be charged after the trial period ends, with full access to our content based on the plan you choose.
After your 3-day free trial you gain access to the full breadth of our content, depending on the subscription plan you choose. We offer free trials for our monthly, bi-annual and annual plans, and each plan provides a different scope of access and benefits.
If your employer has agreed to fund your learning, in part or in full, please contact us at info@lunartech.ai. We’ll work with you and your employer to set up everything needed for your subscription.
Yes. Our support remains strong both during your free trial and after you subscribe. Reach the team at info@lunartech.ai with any question, whether you’re onboarding, navigating your trial, or need help as a member. Subscribers to the yearly plan also receive dedicated support in one-on-one sessions.
You can upgrade your plan at any time during your subscription. Downgrading is not available for bi-annual or annual plans, but you can cancel and then choose a new plan that suits you better. For bi-annual and annual plans we also offer payment in instalments; please reach out to learn more.
The course
Build the skills behind NCP-AIN.
Prove you can build and run networks for AI clusters. Start with a free trial, or enroll in a plan and work through the course at your own pace. NVIDIA runs the exam and issues the credential.
LUNARTECH is independent of the vendors and certifying bodies listed here. Their names, exam titles and marks belong to them, exams are run by them, and credentials are issued by them. Reviewed 5 November 2026. Catalogues change often, so confirm the details on the official page before you book.
