Skip to content

NCP-AIN: mapping the blueprint and closing the gaps

S4·E5Who answers at 02:10 · The sorting hall, the morning the first racks are energised

S4·E5Analyze~25 minsources checked todaylab mutates hardwareverified against NCP-AIN NCP-IB NCA-AIIO NCP-ARI NCP-AII and the NVIDIA certification catalog page all re-fetched 2026-09-09; weights unchanged from the 2026-09-07 research fetch

Builds on: The Ethernet vs InfiniBand decision framework, The networks in an AI factory, and what the RA refuses to specify

Before you read: what do you already know?

3 quick questions. Wrong answers are fine and expected; trying first makes the lesson stick.

After this lesson you can

  • State the NCP-AIN exam parameters and the six published domains with their weights.
  • Map each published objective bullet to a lesson on this platform and classify it as covered partial or gap.
  • Allocate study hours in proportion to the published weights instead of to the domains you already enjoy.
  • Distinguish a gap that reading closes from a gap that only hardware or simulator time closes.
  • Place NCP-AIN among the neighbouring NVIDIA exams including the retired NCP-IB.

Episode 5 — Who answers at 02:10

The situation · The sorting hall, the morning the first racks are energised

The night-shift operator asks it while labelling a cable tray, which is when he asks everything: at 02:10, when the fabric misbehaves, who does he call, and what is that person actually qualified to do? Procurement is standing behind him wanting the same answer, but as a line in a support contract with a lead time attached.

“We will read up” is not an answer, and NVIDIA already publishes the instrument that is. NCP-AIN is 70 to 75 questions in 120 minutes, valid “for two years from issuance”, and its stated prerequisite is “Two to three years of operational experience working in a data center with NVIDIA hardware solutions”.[1] That sentence is the entire design constraint on a study plan: it is an operations exam, so an hour spent reading is worth less than an hour spent running a command and watching it fail.

Certifications exist to make a routing decision legible to people who cannot audit your skills directly — which calls come to you, and which go to a specialist. What makes this one useful to you personally is the published weighting: six domains, with Spectrum Networking at 30% and InfiniBand Networking at another 30%.[1] Sixty per cent of the exam lives in two protocol domains, and only one of them is the one you touch daily in the Dell lab.

A certification is a routing decision made legible, not a badge — so build the map, date it, and hand it over.

What he gets is not a badge. It is one page.

1The blueprint as published

NCP-AIN is 70 to 75 questions in 120 minutes for $400 and the certification is “valid for two years from issuance”.[1] The published prerequisite is the sentence that should set your expectations: “Two to three years of operational experience working in a data center with NVIDIA hardware solutions. The candidate should be able to deploy and manage NVIDIA AI networking infrastructure in support of AI workloads.”[1] That is an operations exam. A study plan built entirely from reading is planning against the wrong instrument.

Six domains, with the weights exactly as NVIDIA publishes them:[1]

Domain Weight
AI Data Center Design and Optimization 5%
NVIDIA Spectrum Networking 30%
NVIDIA InfiniBand Networking 30%
Kubernetes Integration 5%
Troubleshooting Tools 20%
Automation and Configuration 10%

Every figure above was re-fetched from nvidia.com on 2026-09-09 and matched the 2026-09-07 research fetch bullet for bullet.[1] Record the date you fetched it next to the table you build, because that date is the only thing that tells a future reader whether your plan is still valid.

ProductSpeedPCIeRoleGPU generation
NIC
SuperNIC (no Arm)
SuperNIC (no Arm)
DPU
DPU
SuperNIC (Arm inactive)
DPU / storage processor
Ethernet switch
Ethernet switch
InfiniBand switch
InfiniBand switch

⚠ = not confirmed on a fetched primary source (hover for why). Facts as of DOCA 3.5.0 (Sep 2026). Selections are saved.

Filter by the hardware each domain actually needs. Note which rows your Dell lab reaches with a BlueField-3 or ConnectX host and which rows need a Spectrum switch a simulator or an InfiniBand fabric.

2Weight is an hour budget not a table of contents

Convert the weights into the two numbers that drive a plan. At the midpoint of the published range - 72 questions - a 30% domain is about 22 questions and the 5% design domain is about 4.[1] Both figures are derived by multiplication and NVIDIA does not print a per-domain question count, so label them derived wherever you write them down. Time per question at 120 minutes and 72 questions is about 1.7 minutes, also derived.[1]

The uncomfortable arithmetic for a Dell lab engineer: the two domains you can practise daily on your own hardware - Spectrum-side RoCE and SuperNIC work through BlueField-3 and ConnectX - sit inside a 30% domain you are already strong in, while the 30% InfiniBand domain covers objects you have never operated: initial provisioning with high availability, partition keys for multi-tenancy, QoS and adaptive routing, and UFM for link status and bandwidth monitoring.[1][9] Proportional allocation therefore sends your largest block of hours to the material you have the least equipment for. That is the correct answer and it is why people quietly do the opposite.

The design material this course teaches is 5%.[1] Four modules of reference-architecture work buy you roughly four questions. That is not an argument against the course - the RA material is what you get paid for as an FAE - but it is an argument against treating the course you enjoyed as a proxy for exam readiness.

3Mapping objectives to lessons

Do the map at the granularity NVIDIA publishes it: one row per objective bullet, not one row per domain. The Spectrum domain alone has eight bullets and marking it “covered” because you know RoCE hides the seven you do not.[1]

Worked map for this platform, by domain:

  • AI Data Center Design and Optimization (5%) - covered. The three bullets are AI factory architecture and its components, rail-optimized topologies, and GPU-to-GPU communications.[1] They map to ra-m1-01-which-ra-governs, ra-m1-04-the-six-networks, ra-m2-01-rail-optimized and ra-m1-03-scaleup-vs-scaleout.
  • NVIDIA Spectrum Networking (30%) - covered elsewhere. RoCE configuration, QoS/ECN/PFC, adaptive routing, telemetry, BGP-EVPN multi-tenancy, NVIDIA Air, WJH, NetQ, DOCA install and SuperNIC configuration are the spectrumx/ and roce/ tracks plus the DOCA modules.[1] This course touches it only at the design layer - the switch catalog, and the constraint that Spectrum-4 adaptive routing runs at 400G and 200G but not at 800G.[11]
  • NVIDIA InfiniBand Networking (30%) - covered by the infiniband/ track, not by this course.[1][9]
  • Kubernetes Integration (5%) - the k8snet/ track and the DPF material.[1]
  • Troubleshooting Tools (20%) - split. This course contributes only the escalation checklist and support-path triage; cl-resource-query, WJH event analysis, UFM system health and the ib_* command set live in the spectrumx/ and infiniband/ tracks.[1]
  • Automation and Configuration (10%) - NVUE templates and Ansible in the spectrumx/ track.[1]

4The eight documented gaps and which of them are still open

The research notes for this course list eight gaps recorded when they were written: BGP-EVPN multi-tenancy on Cumulus; NVIDIA Air as a first-class lab environment; NetQ operation including flow telemetry, WJH triage and histogram reading; cl-resource-query and Spectrum-X resource triage; the entire InfiniBand 30% domain; the ib_write_lat / ib_write_bw / ibping / ibstat / ibdiagnet / ibnodes / iblinkinfo command set; NVUE templates and Ansible playbooks; and SHARP with sharp_am, libsharp and IPoIB TCP port 6126.[1][10]

Re-run that list against the platform’s current lesson catalog and every one of the eight now has a lesson behind it - spectrumx-m2-05-evpn-tenancy, spectrumx-m5-01-wjh, spectrumx-m5-02-netq, spectrumx-m6-01-triage-toolkit, spectrumx-m5-04-nvue-automation, the whole infiniband/ track, infiniband-m4-03-ibdiagnet and infiniband-m3-03-sharp. That is the analysis this lesson wants you to perform yourself rather than accept: a gap list has a date on it and decays.

What has not changed is the second column. A lesson closes a reading gap. It does not close a hands-on gap, and NCP-AIN is written for someone with operational experience.[1] Sort the remainder honestly:

  • Closable in simulation. DSX Air simulates Spectrum-X switches, ConnectX SuperNICs and BlueField DPUs at “hundreds of switches and thousands of GPUs”, integrates NetQ, and offers a free trial.[7] NVIDIA Air also builds digital twins of the switch systems including “logical instances of every switch and cable” for validating automation and upgrade procedures.[8] Air is not a workaround here - it is one of the published objective bullets in the 30% Spectrum domain, so simulator time is examinable time.[1]
  • Closable on the Dell lab. Everything NIC-side: DOCA-Host install, SuperNIC configuration, RoCE counters, firmware queries on the BlueField-3 and ConnectX hosts.[1][12]
  • Not closable without borrowed hardware. InfiniBand fabric bring-up, subnet-manager high availability, PKeys, UFM system health, and physical-layer behaviour such as optics faults and PFC storms on real switch buffers.[9][7] Book customer-lab or bench time for these or accept that roughly a third of the exam rests on reading alone.
Build one domain's plan

Domain: Troubleshooting Tools, 20%. Published bullets: cl-resource-query for resource allocation in Spectrum-X environments; WJH services for real-time event analysis; verify low-latency interconnects between GPUs CPUs and storage; UFM system health for InfiniBand; and the command set ib_write_lat, ib_write_bw, ibping, ibstat, ibdiagnet, ibnodes, iblinkinfo.[1]

Five bullets. Derived question estimate: 20% of 72 is about 14 questions, so roughly 2.9 questions per bullet - derived, not published.[1]

Status per bullet against the catalog and against Franco’s hardware: cl-resource-query is reading-covered by spectrumx-m6-01-triage-toolkit and hands-on gap - needs a Spectrum switch or Air. WJH is reading-covered by spectrumx-m5-01-wjh and reachable in Air, which produces the telemetry integrations named on the Air page.[7] Low-latency interconnect verification is partly reachable on the Dell lab with the ConnectX host. UFM system health is reading-covered by the infiniband/ track and a hard hardware gap - UFM ships as containers or dedicated appliances against a real fabric.[9] The ib_* set is reading-covered by infiniband-m4-03-ibdiagnet and a hard hardware gap; ibstat alone runs on the local BlueField-3.

Hour budget for a 60-hour plan: 20% of 60 is 12 hours - derived. Split 4 hours reading, 5 hours in Air, 3 hours on the Dell lab, and flag the UFM and ibdiagnet bullets as needing a booked customer-lab session that no amount of the 12 hours replaces.

5Where NCP-AIN sits among the neighbouring exams

Four facts to keep straight, all re-fetched on 2026-09-09.

NCP-IB is retired. The page states: “This certification exam has been retired. Please refer to our new Professional AI Networking Exam.”[2] Anyone planning an InfiniBand-only certification path is planning against a dead exam, and the reason it matters is structural: NCP-AIN absorbed InfiniBand as a 30% domain, so the InfiniBand skills did not become less valuable - they became half of a broader exam.[1][2]

NCA-AIIO is the associate step. 50 questions, 60 minutes, $125, valid two years, prerequisite “a basic understanding of data center infrastructure”, weighted Essential AI Knowledge 38%, AI Infrastructure 40%, AI Operations 22%.[3] This course covers most of the 40% AI Infrastructure domain - cluster components, power and cooling, networking for AI, high-speed networks and DPU benefits.[3] The other 60% is out of scope here.

NCP-ARI and NCP-AII are adjacent professionals not substitutes. The NVIDIA certification catalog lists both alongside NCP-AIN at professional level.[4] NCP-ARI (AI Rack and Interconnect) is 70 questions, two hours, $400, with its heaviest domain High-Density Cabling Installation for AI Clusters at 30%, plus pre-deployment planning and site assessment 11%, rack infrastructure preparation 10%, cable support systems and weight management 7%, testing verification and documentation 12%, safety standards and compliance 10%, and AI infrastructure basics 20%.[5] NCP-AII (AI Infrastructure) is 70-75 questions, 120 minutes, $400, weighted Cluster Test and Verification 33%, System and Server Bring-up 31%, Control Plane Installation and Configuration 19%, Troubleshoot and Optimize 12%, Physical Layer Management 5%.[6] NCP-ARI overlaps this course’s optics, cabling and power material more than NCP-AIN does - worth knowing when a Dell account’s real question is about a cabling plan rather than a fabric design.

Qualification request · decision 1/6A 0 · P 0 · S 0

Brief — Email, Tuesday: "We want to offer the BlueField-3 B3140H SuperNIC in the XE9780 for Spectrum-X customers. Can NVIDIA qualify it? I need a plan by Friday."

Dell platform PM (17G AI servers): So — can you get it qualified? What do you need from us to start?

A qualification call where the customer asks what your team is certified to support. Practise answering with the domain map and a date rather than with a badge - and watch what the scoring does to the answer that overstates coverage.

Episode 5 — Case closed

How it ended

The page goes to the operator and to procurement: six domains, an hour budget derived from the published weights and labelled derived, one column for what a lesson closes and one for what only a simulator, the bench or a borrowed fabric closes.[1][7] The InfiniBand row is the largest and emptiest: a subnet manager, a partition key and UFM are objects you have never operated.[1] The cabling row gets a name that is not yours: NVIDIA weights high-density cabling installation at 30% of a separate professional exam.[5]

The board signed on Friday. The label maker is already printing STORAGE BACK END — DO NOT PATCH. And the sentence that closed the case: “Here is what we cover, here is what we do not, and here is the date I re-check.”

Lab

Mutating steps ahead. Needs a maintenance window, out-of-band access (BMC/iDRAC/rshim console) and a rollback path. Record the pre-flight inventory before changing anything. Never on a production host.

Pre-flight inventory (read-only, do this first). On the Dell-lab host record: mlxfwmanager --query output for every adapter, ibstat output, lspci | grep -i mellanox, the installed DOCA-Host version, and the kernel version. Save this file - it is your rollback reference and your evidence of the starting state.

  1. Read-only baseline. ibstat on the BlueField-3 and ethtool -i <iface> on the ConnectX host. Expected: firmware and driver versions that you can compare against a validated-stack row. If ibstat is absent the OFED/DOCA userspace is not installed - that is step 2’s job.
  2. Install DOCA-Host (MUTATING). Follow the DOCA install procedure for your distribution. Rollback: record the exact package set and version before you start; if the install breaks the host, remove the DOCA-Host meta-package and reinstall the previously recorded driver version, then reboot and re-run the pre-flight inventory until it matches the saved file. Do not attempt this on a host anyone else is using.
  3. Verify (read-only). mlxfwmanager --query and ibstat. Expected: adapter PSID, current firmware version, and per-port state. This pair of commands is the concrete evidence behind the blueprint’s DOCA install objective.
  4. Do not flash firmware for this lab. If mlxfwmanager reports available firmware, note it and stop. Firmware update is a separate change with its own rollback plan and is not required by any objective you are closing here.
  5. SuperNIC-side objectives on the ConnectX host (read-only). Record negotiated link speed with ethtool <iface> and RoCE-relevant counters with ethtool -S <iface> | grep -i -E 'pause|ecn|cnp'. Expected: counters present and mostly zero on an idle link. Mark the corresponding blueprint rows dell-lab rather than gap.
  6. Mark the InfiniBand rows honestly. ibstat on a BlueField-3 in Ethernet mode is not InfiniBand fabric experience. Leave the four InfiniBand objective bullets in the borrowed-hardware column.
  7. Optional (customer lab). Book time on an InfiniBand fabric and run ibdiagnet, iblinkinfo, ibnodes and a UFM system-health view. These have no local substitute - the fabric objects they report on do not exist on your bench.

Retrieval check

10 questions from memory. Answer before looking anything up; misses become flashcards.

Explain it to a Dell SE

Explain to your manager in five sentences why your NCP-AIN study plan spends more hours on InfiniBand than on the Ethernet material you use every day.

14 flashcards for this lesson — 0 in deck. Spaced review lives at /review.

Sources

Facts in this lesson were checked against NCP-AIN NCP-IB NCA-AIIO NCP-ARI NCP-AII and the NVIDIA certification catalog page all re-fetched 2026-09-09; weights unchanged from the 2026-09-07 research fetch. Dates are when each page was fetched.

  1. Professional AI Networking Certification (NCP-AIN) | NVIDIA · fetched 2026-09-09
  2. InfiniBand Certification (NCP-IB) | NVIDIA - retired · fetched 2026-09-09
  3. AI Infrastructure and Operations (NCA-AIIO) Certification | NVIDIA · fetched 2026-09-09
  4. NVIDIA Certification catalog · fetched 2026-09-09
  5. AI Rack and Interconnect Certification (NCP-ARI) | NVIDIA · fetched 2026-09-09
  6. AI Infrastructure Certification (NCP-AII) | NVIDIA · fetched 2026-09-09
  7. NVIDIA DSX Air Platform for AI Factory Simulation · fetched 2026-09-07
  8. NVIDIA Spectrum SN5600 Series Switches Datasheet (Dell-branded) · fetched 2026-09-07
  9. NVIDIA UFM - Unified Fabric Manager · fetched 2026-09-07
  10. NVIDIA SHARP Installation · fetched 2026-09-07
  11. Equal Cost Multipath Load Sharing incl. Adaptive Routing | Cumulus Linux 5.13 · fetched 2026-09-07
  12. NVIDIA Spectrum-X Validated Solution Stack · fetched 2026-09-07

The same idea elsewhere

Other lessons that cover this ground, sometimes from another course's angle.