I85PREP
⌘ K
NVIDIA NCP-AII study materials
NVIDIA NCP-AII study resources

NVIDIA NCP-AII: System and Server Hardware Installation

NVIDIA NCP-AII · Work through your exam, one topic at a time.

All guide sections

System and Server Hardware Installation

What this domain tests

Turning a bare chassis into a GPU server the software stack can see and trust: PCIe slot and lane planning, SXM/HGX baseboard handling, power delivery, airflow, and the firmware settings that decide whether accelerators enumerate at all.

Core rules

  • Enumerate before you diagnose software. lspci -d 10de: -nn shows what the BIOS exposed over PCIe; nvidia-smi shows what the driver bound to. Those are two different failures with two different fixes.
  • Slot choice matters. A GPU in a CPU-attached x16 slot is not equivalent to one behind a PCIe switch. Confirm NUMA locality with lscpu -e and nvidia-smi topo -m, and enable Above 4G Decoding, Resizable BAR, and the highest supported link speed in firmware.
  • HGX/SXM baseboards are factory-sealed units. Eight SXM GPUs and the NVSwitch fabric ship as one assembly. Do not reseat them; instead ensure nvidia-fabricmanager runs on the host, because NVSwitch is unusable without it.
  • Power and airflow are installation tasks, not afterthoughts. SXM parts draw 400-700 W each. Feed the baseboard from separate PSU rails, never daisy-chain PCIe power cables, and match chassis fan direction to the GPU heatsink airflow.

Worked example

A technician installs eight GPUs, but nvidia-smi lists four. Correct sequence: run lspci -d 10de: -nn first. If only four NVIDIA controllers appear, the fault is physical or firmware - reseat the card, reseat or swap the riser, enable Above 4G Decoding, and verify PSU headroom. If all eight appear in lspci but four in nvidia-smi, hardware is fine: move to driver binding or Fabric Manager. The reasoning is that PCIe enumeration is a BIOS/hardware function and driver binding is an OS function, so the two commands partition the problem.

Traps that cost marks

Assuming a dead GPU when a slot is bifurcated x8/x8 and only one riser is populated; blaming NVSwitch for a stopped Fabric Manager; ignoring NUMA placement and later wondering why host-to-device copies are slow.