NVIDIA-Certified Professional: AI Infrastructure (NCP-AII)

The NCP-AII certification validates advanced skills in designing, deploying, and optimizing AI infrastructure powered by NVIDIA GPUs. Unlike model-focused certifications (like LLM or Agentic AI tracks), this one concentrates on the infrastructure layer that enables AI workloads at scale — including compute, networking, storage, and orchestration. It is a professional-level certification intended for infrastructure and platform engineers.

---------- Question 1

To facilitate the use of various AI models and tools, an administrator needs to install the NGC CLI on the cluster's hosts. What is the main benefit of using the NGC CLI in a professional AI infrastructure, and how does it integrate with the control plane's workflow?

1. It allows users to download and manage optimized AI containers, pre-trained models, and scripts directly from the NVIDIA GPU Cloud repository.
2. It provides a graphical user interface for monitoring the temperature of individual GPU cores across the entire cluster in real-time.
3. It replaces the standard Linux shell with a specialized AI-focused terminal that only supports Python-based commands and libraries.
4. It is used to physically format the hard drives of the compute nodes before the operating system is installed by the Base Command Manager.

---------- Question 2

A system administrator identifies a faulty H100 GPU in an HGX baseboard that is causing consistent Bus Errors. After confirming the hardware failure, what is the correct high-level procedure for replacing the GPU while maintaining the integrity of the remaining system components?

1. Power down the system, follow anti-static procedures, remove the HGX heat sink assembly, replace the specific GPU module applying correct torque to the fasteners, and then re-validate with a burn-in test.
2. The H100 GPUs are hot-swappable; the administrator should use a specialized extraction tool to pull the GPU while the system is running and insert a new one immediately.
3. Use the nvidia-smi -r command to logically reset the GPU, which physically ejects the faulty silicon from the socket so it can be picked up from the bottom of the chassis.
4. Replace the entire motherboard because individual GPUs on an HGX baseboard are permanently soldered and cannot be replaced without specialized factory wave-soldering equipment.

---------- Question 3

When setting up a MIG (Multi-Instance GPU) configuration for a multi-tenant AI environment, an administrator needs to ensure that memory and cache isolation are strictly enforced between different users. Which MIG profile characteristic ensures that compute resources are dedicated and not shared with other instances on the physical GPU?

1. The use of Shared profiles which allow multiple instances to burst into the same L2 cache bank for performance.
2. The selection of a specific Slice (such as 1g.10gb) which provides hardware-level isolation of the memory controller and SMs.
3. Configuring the GPU in Time-Slice mode which uses the OS scheduler to rotate access to the entire GPU every 10ms.
4. Enabling the Overcommit flag in the NVIDIA driver to allow instances to use more than their allocated VRAM when available.

---------- Question 4

A developer needs to run a specialized AI model using a specific version of the NVIDIA Container Toolkit and a custom Docker image. Which command sequence correctly demonstrates how to utilize a GPU within a Docker container on a properly configured NVIDIA-certified host?

1. docker run --gpus all --rm nvidia/cuda:12.0-base nvidia-smi
2. docker run --use-cuda-cores=max -it ubuntu:latest run-ai-model
3. nvidia-docker start --container-id=auto --memory-limit=unlimited
4. docker exec -u gpu-user my_container ./start_training.sh --with-gpu

---------- Question 5

When configuring a BlueField-3 DPU to support AI workloads, which feature must be correctly implemented to allow for efficient communication between the GPU memory and the network without involving the host CPU's system memory?

1. GPUDirect RDMA, which requires the peer-to-peer (P2P) capability to be supported and enabled between the DPU and the GPU over the PCIe bus.
2. NVIDIA SMI migration, which automatically moves the GPU memory pages to the BlueField's internal DDR5 memory during periods of high network congestion.
3. The Slurm scheduler's DPU-plugin, which partitions the DPU's ARM cores into MIG-like instances to handle concurrent SSH connections from cluster users.
4. Encapsulated Remote Port Mirroring (ERSPAN), which allows the BlueField DPU to copy GPU memory directly to the storage fabric for real-time backup.

---------- Question 6

A data center engineer is connecting several NVIDIA DGX nodes to a leaf-and-spine network fabric. To ensure optimal performance and avoid signal degradation, the engineer must validate the cabling and transceivers. Which specific action correctly identifies a fault in the physical layer when the link fails to come up at the expected 400Gbps or 800Gbps speed?

1. Inspect the optical fiber end-faces for contamination using a digital microscope and verify that the transceiver power levels in the BMC fall within the specified dBm range.
2. Increase the MTU size to 9000 on the BMC management port to force the InfiniBand transceivers to negotiate a higher clock speed via the OOB network.
3. Reinstall the NVIDIA Container Toolkit to ensure that the drivers can properly communicate with the physical layer transceivers and reset the link state.
4. Swap the OSFP transceivers with SFP+ equivalents to determine if the high-speed signaling is being restricted by the HGX firmware power limits.
5. None of the above.

---------- Question 7

A cluster node is reporting a 'GPU Fallen Off Bus' error in the system logs. After verifying the physical seating and power connections, what is the next logical step an administrator should take to troubleshoot this hardware fault on an NVIDIA HGX system?

1. Check the dmesg output for PCIe AER (Advanced Error Reporting) messages and use the BMC to check for any critical hardware events or power faults related to that GPU slot.
2. Re-install the NGC CLI on the head node and use the 'ngc fix-gpu' command to remotely reset the PCIe bus registers on the affected compute node.
3. Increase the fan speed to 100% via the BIOS and then disable the MIG configuration to allow the GPU to re-sync its clock with the BlueField-3 DPU.
4. Swap the InfiniBand transceivers between the faulty node and a known-good node to see if the GPU error follows the network cable.

---------- Question 8

An engineer is using the High-Performance Linpack (HPL) benchmark to validate a single node's compute performance. If the HPL results are inconsistent across multiple runs, what is the first hardware-related parameter that should be monitored via the NVIDIA SMI tool during the test?

1. GPU temperature and power draw to check for thermal throttling or power limit capping that could be causing performance fluctuations.
2. The number of active SSH sessions to the BlueField-3 DPU to ensure that the ARM cores are not being overwhelmed by management traffic.
3. The firmware version of the local SATA boot drive to ensure it is compatible with the version of the NVIDIA Container Toolkit being used for HPL.
4. The light levels of the OOB management transceivers to ensure that the BMC is not losing connectivity with the Base Command Manager head node.

---------- Question 9

When performing the initial bring-up of an NVIDIA HGX H100 system, an administrator notices that the Baseboard Management Controller (BMC) reports a power capping event during the hardware validation phase. Given that the rack power distribution units (PDUs) are within limits, which specific step should be prioritized to ensure the GPU-based server meets the power requirements for high-performance AI workloads?

1. Decrease the GPU clock frequency using nvidia-smi to stay under the current power threshold.
2. Verify the Power Supply Unit (PSU) redundancy policy in the BMC and ensure all power cables are seated and connected to independent circuits.
3. Update the TPM firmware to version 2.0 to allow for higher power draw authorization from the motherboard.
4. Reinstall the NVIDIA Container Toolkit to recalibrate the power sensing logic of the operating system.

---------- Question 10

During the final verification phase of an AI factory deployment, the team executes a High-Performance Linpack (HPL) test. The results show a significant Rmax value drop compared to the Rpeak theoretical performance. Which cluster-level assessment tool is best suited for identifying if the issue is a specific limping node or a general network congestion issue?

1. ClusterKit; it performs multifaceted node assessments and identifies outliers in performance across the entire cluster.
2. The DOCA Benchmarking tool; it isolates the DPU performance from the GPU performance to check for CPU bottlenecks.
3. The Slurm squeue command; it identifies which jobs are pending and allows the administrator to prioritize the HPL task.
4. The ping utility; it checks for basic ICMP connectivity between the head node and the compute nodes.

---

Are they useful?

Click here to get 360 more questions to pass this certification at the first try! Explanation for each option is included!

Follow the below LINKEDIN channel to stay updated about 89+ exams!

AI Strategy and Engineering Hub