NCP-AII Tests Bringing a GPU Cluster Up and Proving It Works — Not Buying One
A hands-on professional exam for engineers who rack, image, test and repair DGX and HGX clusters. Think in deployment order: firmware baseline before any test, single-node before multi-node, evidence before swapping parts.
Check Your Readiness →Most candidates understand NVIDIA-Certified Professional: AI Infrastructure (NCP-AII) concepts — and still fail. This exam tests how you apply knowledge under pressure.
Questions are operational: a stage of a DGX or HGX rollout, an nccl-tests table, an HPL summary or an XID line, and you choose the next step or the faulty layer. Testing and bring-up together make up 64%, so sequencing and output reading decide the result.
Investigate the InfiniBand compute fabric and cabling
A single-node NCCL test never touches the network; traffic runs over NVLink and NVSwitch. Check nvidia-smi nvlink link state and speed, and confirm Fabric Manager is running at the driver's version.
Reinstall the GPU driver and reboot the node
If the BMC, which works without the host OS, also misses the device, the cause is hardware, seating, power or firmware. Check the SEL and XID logs, run nvidia-bug-report.sh, drain the node, then locate the GPU by bus ID rather than index.
Install the GPU driver inside the container image
The NVIDIA Container Toolkit mounts the host driver into containers; a driver in the image conflicts with the kernel module. Configure and restart the runtime, then pass --gpus. Under Slurm, Pyxis adds --container-image to srun but needs Enroot installed.
Record the TFLOPS figure as the node's acceptance result
A failed residual means the computation was wrong, so the run is invalid regardless of speed. Investigate the node, retune N, NB and the P x Q grid within memory, and rerun before burn-in.
Replace the PSUs, since idle power readings already passed validation
Peak draw appears only under full GPU load, and an eight-GPU server can draw around 10 kW. Plan density by rack kilowatts and cooling, not rack units, and validate power under stress.
NVIDIA validates BMC, BIOS, GPU, NVSwitch and NIC firmware as one set. Updating one component from a release note creates an untested combination, and spare parts may ship with other versions, so align firmware before acceptance tests and after any replacement.
algbw depends on the collective; busbw is normalized so it can be compared with hardware bandwidth. Run all_reduce_perf on one node first, which exercises NVLink and NVSwitch only, then multi-node across the rails.
Fabric Manager must run the same version as the GPU driver, or the NVSwitch fabric does not initialize and multi-GPU work fails. The DOCA host stack for ConnectX and BlueField is a separate install from the GPU driver.
BCM reprovisions nodes from their software image, so manual edits vanish. Change the node category or the software image instead. BCM HA counts as verified only after status checks pass and a controlled head-node failover succeeds.
MIG mode takes effect only after a GPU reset or reboot. Profiles are fixed sizes such as 1g.10gb, 3g.40gb and 7g.80gb on an 80 GB GPU, up to seven instances. Time-slicing shares a GPU without memory or fault isolation.
Strengthen weak areas with exam-style practice questions and detailed explanations.
Simulate the real exam experience and assess your readiness under timed conditions.
Review key concepts, objectives, and exam topics in one place.
Get personalized explanations, learning recommendations, and instant answers.
Follow a structured learning path designed to help you prepare efficiently.
NCP-AII rewards doing cluster bring-up steps in the right order. Find out where your sequencing slips.