This is the condensed review layer for NCP-ARI (NVIDIA-Certified Professional, AI Rack and Interconnect): the cabling decision rules, prep checklists, and validation workflow that the exam's 70 questions keep circling back to. Because NCP-ARI tests the physical build of AI clusters, this sheet reads like a foreman's pocket card rather than a command reference. Use it for final-week passes and as a fluency check, because every rule here should come out of your mouth without a lookup. Scenario questions reward technicians who can decide fast. Depth lives in the domains breakdown, and logistics live in the complete guide.
Exam Quick Facts
Cabling Decision Rules (High-Density Cabling, 30%)
The heaviest domain is about choosing the right link medium and then physically managing thousands of them. InfiniBand NDR runs 400 Gb/s per port and XDR runs 800 Gb/s, both landing in OSFP (octal small form factor pluggable) cages, and the same three media classes answer nearly every reach question the exam can ask.
DAC vs ACC vs Fiber by Reach
| Media | Reach class | Choose it when | Watch out for |
|---|---|---|---|
| Passive DAC (direct attach copper) | In-rack and adjacent-rack runs up to roughly 3 m | Both ports live in the same rack or the next one over, and you want zero added power draw and minimal latency | This is the thickest and stiffest cable class, so bundle weight and bend radius drive the routing plan |
| ACC (active copper cable) | Mid-reach runs up to roughly 5 m | The run slightly outreaches passive copper, but a transceiver pair would cost more power and money than the link needs | Active electronics in each connector end draw power and add one more thing that can fail |
| Multimode fiber with transceivers | Row-scale runs up to about 50 m | Leaf-to-spine links inside a pod sit beyond any copper reach | Every connector endface is a contamination risk, so inspect and clean before mating |
| Single-mode fiber with transceivers | Long runs of 100 m and beyond | The link crosses rows, pods, or rooms | This is the highest transceiver cost per link, and APC polish and connector keying must match the plan exactly |
MPO/APC handling rules
An MPO (multi-fiber push-on) connector packs a dozen or more fibers behind a single rectangular ferrule, which is how one plug can carry a 400 or 800 gigabit port. APC (angled physical contact) polish cuts that ferrule endface at 8 degrees, so stray reflections scatter into the cladding instead of bouncing straight back toward the laser. The handling rules follow from those two facts.
- Dust caps stay on until the moment of mating, and they go straight back on whenever a connector comes out of a port.
- Never touch a ferrule endface, because skin oil scatters light as effectively as dust does.
- The discipline is inspect, clean, inspect: scope the endface, clean with a dry one-click tool only if the scope shows contamination, then scope again before mating.
- APC mates only with APC. The green housing is the field marker, and forcing an angled connector into a flat-polish coupling leaves an air gap that destroys the link budget.
- Respect the key orientation on every MPO plug, because polarity mistakes made here surface later as validation mismatches.
Most dead optical links are dirty links
Contaminated endfaces are the most common physical cause of failed or error-prone optical links, which is why the inspect, clean, inspect habit shows up in both the cabling domain and the testing domain. If a fiber link fails validation, assume contamination before you assume damage.
The 50/50 left-right routing rule
Split every rack's cable population evenly between the left and right vertical management channels. The rule exists because a single overloaded side blocks the airflow boundary, buries the OSFP cages you will need to reach during service, exceeds the weight one channel can safely carry, and forces longer worst-case runs. An even split keeps both channels inside their fill limits and keeps every port reachable without disturbing a dressed bundle.
The 30 mm bend radius
Treat 30 mm as the minimum bend radius for the high-density cable plant. A fiber bent tighter leaks light, which appears as attenuation and a rising BER (bit error rate, the fraction of transmitted bits that arrive corrupted). A copper DAC bent tighter deforms its internal geometry and degrades signal integrity, so the link may train at a lower speed or width, flap under load, or fail validation outright. Dress bundles in wide sweeps, use hook-and-loop straps in place of cinched zip ties, and never let a cable turn a hard 90-degree corner.
Preparing for NCP-ARI? Practice with 455+ exam questions
What Makes an AI Rack Different (AI Infrastructure Basics, 20%)
| Dimension | Conventional enterprise rack | AI cluster rack |
|---|---|---|
| Power density | A typical rack draws roughly 5 to 15 kW. | A dense AI rack can pass 100 kW, which is why direct busbar power delivery replaces cords and rack PDUs. |
| Cooling | Air cooling with hot and cold aisle separation handles the load. | Direct-to-chip liquid cooling carries most of the heat, and air handles only the remainder. |
| Networking | A few uplinks per rack are enough. | Every GPU gets its own fabric port, so a single rack can terminate hundreds of high-speed links. |
| Form factor | Standard-width frames dominate. | Extra-wide frames make room for coolant manifolds, busbars, and dense cable channels. |
| Building block | Racks deploy one at a time. | SuperPOD designs deploy scalable units, groups of racks cabled together as one fabric, so a miscabled rack degrades an entire unit. |
The fabric behind all those ports is usually a fat tree, a layered switch topology that preserves full bandwidth between any two GPUs in the cluster, and it is often rail-optimized, meaning GPU port 1 on every server lands on the same leaf switch. You build the physical half of that design, and the network team configures the logical half; the NCP-AIN networking guide covers that upper layer if you want the context.
CVT Workflow and Failure Causes (Testing and Verification, 12%)
The Cable Validation Tool (CVT) is NVIDIA's checker for one question: does the cabling you actually built match the cabling the design intended? NVIDIA offers a free self-paced CVT Fundamentals course, and the exam expects you to know where the tool sits in the workflow and which physical mistakes produce which findings.
- Load the planned topology into CVT so the tool knows what every port should connect to.
- Cable to the plan, labeling both ends of every run as you go.
- Run discovery, which lets CVT read what each switch and adapter port actually sees on the other end.
- Review the mismatch report, which flags missing links, swapped ports, and links that trained below their intended speed or width.
- Fix, rerun, and repeat until the report is clean, then export it into the as-built documentation, the delivered record that the site matches the design.
Common CVT Failure Causes
| Finding | Most likely physical cause | First move |
|---|---|---|
| A link is missing entirely | An unseated cable or transceiver, or a port that was skipped during installation | Reseat both ends and confirm the run against its labels |
| Two ports are swapped | Two cables crossed inside a dense bundle, often as mirror-image errors between the left and right channels | Trace both runs by label, swap them at the port, and rerun discovery |
| A link trains below its intended speed or width | A bend radius violation, damaged conductors, or the wrong cable class for the reach | Redress the run with proper sweeps and check the media class against the plan |
| An optical link shows high errors | A contaminated or scratched ferrule endface | Inspect, clean, inspect, and then remate the connector |
| Fiber polarity or mapping is wrong | MPO key orientation or trunk polarity mistakes | Verify connector keying against the wiring diagram before touching anything else |
Site-Survey Checklist (Pre-Deployment Planning, 11%)
A CDU (coolant distribution unit) is the heat exchanger that isolates the facility water loop from the cleaner loop that actually touches the servers, and its placement is one of the things a survey has to settle. Everything in this table gets verified before hardware ships, because every line item is far cheaper to fix while the floor is still empty.
Site Survey: Verify Before Hardware Ships
| Area | What you verify | The failure it prevents |
|---|---|---|
| Floor load | Static load under each rack footprint plus the rolling load along the entire delivery path | A floor rated for the standing rack that was never checked for the loaded pallet rolling across it |
| Clearances | Door widths, ramp angles, elevator weight ratings, and front and rear service space at the final position | A rack that arrives on site but cannot physically reach its row or be serviced once it is there |
| Power readiness | Feed capacity, breaker sizing, busway routing, and redundancy measured against the full projected draw | Energizing-day surprises when the room cannot actually carry the load it was sold as carrying |
| Cooling readiness | Facility water loop capacity, CDU placement, and airflow for the heat that liquid does not capture | Thermal throttling and emergency shutdowns during the first sustained training run |
| Staging space | A secure area to receive, unbox, and stage components near the installation floor | Crates blocking aisles and expensive components being handled twice |
Master These Concepts with Practice
Our NCP-ARI practice bundle includes:
- 7 full practice exams (455+ questions)
- Detailed explanations for every answer
- Domain-by-domain performance tracking
30-day money-back guarantee
Rack Prep Checklist (Rack Infrastructure Preparation, 10%)
Containment and airflow zones
- Close the containment before you chase airflow numbers: horizontal or vertical containment panels seal the hot and cold aisles into separate zones, and every later thermal measurement assumes that boundary holds.
- Every unused U position gets a blanking panel, every cable penetration gets a brush grommet, and every side gap gets sealed, because any hole in the airflow boundary lets hot exhaust recirculate straight back into the intakes.
Liquid-cooling manifold sequence
- Mount the supply and return manifolds, the vertical pipes that distribute coolant to every server position, and verify their orientation against the drawing.
- Pressure-test and leak-check the loop before any server hose connects, with drip protection already in place.
- Connect server hoses through dry-break quick disconnects, fittings that seal instantly on release so an accidental disconnect drips almost nothing.
- Bleed trapped air from the loop, then verify flow rate and temperature at the CDU.
- Energize compute only after coolant flow is verified, because in a liquid-cooled rack the coolant always moves before the power does.
Busbar rules
A DC busbar is a rigid conductor running the height of the rack that servers tap directly, and it replaces the individual power cords a conventional rack would use.
Treat every busbar as energized
Work near a busbar only after lockout-tagout, the padlock-and-tag procedure that keeps a circuit de-energized while people are on it, has been applied and verified. Protective covers stay on until energization is scheduled, tools used near the bar are insulated, and the torque on every connection gets checked against specification before power-on.
Safety Triggers (Safety, Standards, and Compliance, 10%)
Stop-Work and Escalation Triggers
| You encounter | The rule | Why it exists |
|---|---|---|
| A component above the two-person lift threshold, around 23 kg or 50 lb | Bring a second person or a mechanical server lift, and always use the lift for heavy gear going above chest height | Dropped hardware injures people and destroys six-figure servers in the same second |
| Work above comfortable standing reach | Use a rated ladder or platform every time | Improvised footing such as chairs and crates is a leading cause of installation falls |
| Coolant on the floor or on equipment | Stop, isolate the affected loop, and protect powered electronics before starting cleanup | Liquid meeting an energized busbar is an electrical incident already in progress |
| An exposed conductor or a missing busbar cover | Stop work until the circuit is verified de-energized under lockout-tagout | The procedure exists so that de-energized gets verified rather than assumed |
| Damaged PPE, a frayed cable, or a cracked connector | Tag it out and replace it, because damaged gear never returns to the pool | A known-bad component that stays in circulation will eventually be used |
Weight-Management Rules of Thumb (Cable Support Systems, 7%)
- Keep tray fill at or below roughly half of the tray cross-section, both to stay inside its weight rating and to leave room for the additions that always come later.
- Support bundles at short, regular intervals so no connector ever carries cable weight, because a hanging cable pulls on its port and slowly works itself loose.
- Add strain relief at every transition where a bundle changes direction or drops from a tray into a rack.
- Layer copper low and fiber high, because DACs are the heaviest cables in the plant and stacking them on top of fiber crushes it past its bend limit.
- Plan pathways for total plant weight, because hundreds of copper cables add hundreds of kilograms of load to trays, supports, and the racks themselves.
- Use hook-and-loop straps for dressing, because a cinched zip tie concentrates pressure at one point and creates bend violations you cannot see inside the bundle.
Exam-Format Facts Worth Cold Recall
- The exam runs 70 questions in 120 minutes, which works out to about 100 seconds per question.
- It costs $400 per attempt and is delivered in English through online remote proctoring on the Certiverse platform.
- NVIDIA does not publish a passing score, so calibrate against consistent practice-test performance in the high 70s and above.
- The credential stays valid for 2 years, and you recertify by retaking the current version of the exam.
- High-density cabling (30 percent) plus AI infrastructure basics (20 percent) make up half the exam, so your final week should be weighted the same way.
- Every question is answered at a screen, so the physical procedures in this sheet get tested as scenarios where you pick the next correct step.
Final-Week Usage
Run this sheet top to bottom once a day and mark every rule that produces hesitation, then close each gap the same day with targeted question practice. Preporato's NCP-ARI practice tests mirror the exact 7-domain blueprint with 7 full-length exams (490 questions, every answer explained) framed around a greenfield DGX SuperPOD build, which is the closest a screen gets to a cabling floor. If exam day is still several weeks out, the 6-week study plan sequences this material at a saner pace, and the first-attempt strategy covers pacing and question triage for the day itself.
Final-Week Checklist
0/7 completedSources:
- NVIDIA Certification Program
- NVIDIA DGX SuperPOD Documentation
- NVIDIA LinkX Cables and Transceivers
- NVIDIA Training: Cable Validation Tool (CVT) Fundamentals
Last updated: August 8, 2026
Ready to Pass the NCP-ARI Exam?
Join thousands who passed with Preporato practice tests