The foundational vocabulary of a rack-scale AI system — what's inside a GB200 NVL72, how power, cooling, and interconnect fit together, and the terms every other NCP-ARI domain assumes you know.
Domain 1 is 20% of the exam — second only to cabling — but its real weight is bigger than that: every other domain assumes this mental model. You can't assess a site, prep a rack, land cabling, or verify a link if you can't name the components and explain how power, cooling, and interconnect interact. Get fluent here first.
A modern AI cluster isn't a room of independent servers — it's a small number of very dense, liquid-cooled rack-scale systems wired together into one giant accelerator. NVIDIA's GB200 NVL72 is the reference design this exam is built around, so we'll use it as our worked example throughout.
In one sentence: a GB200 NVL72 packs 72 Blackwell GPUs and 36 Grace CPUs into a single liquid-cooled rack drawing roughly 120 kW, behaving as one unified GPU thanks to a fifth-generation NVLink fabric. Your job as an AI Rack & Interconnect technician is to physically build, power, cool, cable, and validate that rack correctly.
Learn the rack from the trays outward. A standard GB200 NVL72 is built from a repeating stack of trays plus the power and cooling plumbing that serves them.
Each tray holds two GB200 Superchips — so 2 Grace CPUs + 4 Blackwell GPUs per tray. 18 trays × 4 GPUs = 72 GPUs. Liquid-cooled via cold plates on every CPU and GPU.
The scale-up fabric. Nine switch trays connect every GPU to every other GPU in the rack over 5th-gen NVLink at 1.8 TB/s per GPU — the rack acts as one big accelerator.
Power shelves rectify facility AC to DC and feed a vertical busbar at the rear. Trays clip directly onto the busbar instead of using dozens of individual cords.
Pushes cool liquid through in-rack manifolds to the cold plates and returns warm liquid to the facility loop. An in-rack CDU can reject on the order of ~250 kW of heat.
The backbone that wires the compute trays to the switch trays. In NVL72 this is a dense copper cable spine — copper is used inside the rack to save power over short distances.
ConnectX SuperNICs and BlueField DPUs connect this rack to other racks over InfiniBand or Ethernet — the fiber that leaves the rack and heads to the spine switches.
Don't confuse scale-up and scale-out. NVLink (inside the rack, copper spine) is scale-up — it fuses GPUs into one system. InfiniBand/Ethernet (between racks, fiber) is scale-out — it links many racks into a cluster. Different fabrics, different media, different domains of the exam.
Everything you install belongs to one of three layers. Internalize this and the rest of the certification organizes itself.
Facility feed → power shelves (AC→DC) → vertical DC busbar → trays. High current means grounding, bonding, and redundancy (A/B feeds) are non-negotiable. A single NVL72 rack at ~120 kW draws more than an entire traditional rack row.
Air alone can't cool 120 kW. Direct liquid cooling (DLC) runs coolant through cold plates on the chips; the CDU and manifolds move heat to the facility water loop. Rear-door heat exchangers handle residual air heat. Leak detection and dripless quick-disconnects matter here.
Two fabrics: NVLink (scale-up, copper, inside the rack) and InfiniBand/Ethernet (scale-out, fiber, between racks). This is where high-density cabling — the 30% domain — lives.
"Power in, heat out, data across." Power comes up the busbar, heat leaves through the coolant loop, and data moves across the NVLink and network fabrics. Three verbs, three layers.
The most-tested distinction in this domain. Know which fabric does what, which physical medium it uses, and where its boundary is.
| Fabric | Role | Scope | Typical medium | Domain of the exam |
|---|---|---|---|---|
| NVLink (5th gen) | Scale-up: fuse GPUs into one system | Inside the rack | Copper spine / backplane | Rack build & cabling |
| InfiniBand | Scale-out: low-latency cluster fabric | Between racks | Fiber (MPO/MTP), some DAC/AOC | High-density cabling |
| Ethernet (Spectrum-X) | Scale-out: AI-optimized Ethernet fabric | Between racks | Fiber, DAC/AOC | High-density cabling |
Short runs inside the rack use copper (cheaper, lower power, no optical transceiver needed) — that's why the NVLink spine is copper. Longer runs between racks use fiber with optical transceivers, because copper can't carry those speeds that far. You'll meet the connector types (MPO/MTP), transceivers, DACs, and AOCs in depth in Domain 4.
"More fiber is always better" is wrong. Using an optical transceiver where a cheap copper DAC would do wastes power and money; using copper where the distance demands fiber fails the link. Matching the medium to the run length is a core competency.
| Component | What it is | Why it matters to you |
|---|---|---|
| GB200 Superchip | 2 Blackwell GPUs + 1 Grace CPU, joined by NVLink-C2C | The compute building block; two per compute tray |
| Cold plate | Liquid-cooled metal plate on each CPU/GPU | Primary heat removal; must be seated and leak-free |
| CDU | Coolant Distribution Unit | Isolates the rack (technology) loop from the facility loop; sets flow/temperature |
| Manifold | In-rack coolant distribution pipe with drops per tray | Where quick-disconnects mate; leak-prone if mishandled |
| DC busbar | Shared vertical DC power rail at the rack rear | Trays clip on; high current → grounding/LOTO critical |
| ConnectX SuperNIC / BlueField DPU | Scale-out network adapters | Terminate the fiber that leaves the rack |
| Blind-mate connector | Power/data connector that engages as a tray slides home | No per-tray hand-cabling; must seat squarely |
The fastest way to know this material stuck is to be quizzed on it. FlashGenius has full NCP-ARI practice sets with explained answers for every domain.
Practice AI Infrastructure Basics →NCP-ARI-style questions on Domain 1. Pick an answer to see whether it's right and why.
Not sure where your understanding is thin? Pick what's fuzziest and get a targeted study focus.
Tap a card to flip it. Drill these until the answers are automatic.
For educational use only · Not affiliated with or endorsed by NVIDIA Corporation. GB200 NVL72 specifications are drawn from NVIDIA's published materials and can change; always confirm current figures on the official GB200 NVL72 page and the NCP-ARI certification page.
