Free NCP-AII Practice Test Questions 2026

Explanation:

The question tests the correct method to configure NVIDIA vGPU scheduling on a VMware ESXi host so that all GPUs use equal-share scheduling with the default time-slice length. This ensures fair GPU time allocation among vGPUs without favoring any particular virtual machine.

✅ Correct Option: A. esxcli system module parameters set -m nvidia -p "NVreg_RegistryDwords=RmPVMRL=0x01"
This is the official VMware vSphere command for setting the NVIDIA driver registry key.
RmPVMRL=0x01 explicitly enables the equal-share scheduler with the default time-slice length.
The setting applies to all GPUs on the host.
After running the command, a host reboot (or NVIDIA driver reload) is required for the change to take effect.
This configuration provides predictable fairness across vGPUs while retaining the driver’s standard time-slice duration.

❌ Incorrect Option: B. esxcli graphics module parameters set -m nvidia -p "NVreg_RegistryDwords=RmPVMRL=0x01"
There is no esxcli graphics module parameters namespace. The correct command path is esxcli system module parameters. Using the wrong namespace causes the command to fail.

❌ Incorrect Option: C. esxcli system module parameters set -m nvidia -p "NVreg_RegistryDwords=FRL=0x01"
FRL is not a valid NVIDIA registry key for vGPU scheduling. The correct key is RmPVMRL. This command has no effect on scheduling behavior.

❌ Incorrect Option: D. esxcli system module parameters set -m nvidia -p "NVreg_RegistryDwords=RmPVMRL=0x00"
0x00 configures the best-effort scheduler (the default policy). Best-effort prioritizes overall utilization and can allow one vGPU to dominate the GPU, which does not meet the equal-share requirement.

🔧 Reference:
→ Changing Scheduling Behavior for Time-Sliced vGPUs and MIG-Backed, Time-Sliced vGPUs – NVIDIA Virtual GPU Software Documentation
Confirms that on VMware vSphere the command esxcli system module parameters set -m nvidia -p "NVreg_RegistryDwords=RmPVMRL=0x01" sets equal-share scheduling with the default time-slice length for all GPUs.

Explanation:

The question tests how to verify Multi-Instance GPU (MIG) support and configuration on an NVIDIA GPU. The nvidia-smi utility provides GPU management information, including whether MIG mode is enabled. It can display the MIG mode status for supported GPUs.

🟢 Correct Option:

D. nvidia-smi
nvidia-smi is NVIDIA's standard command-line utility for monitoring and managing supported GPUs. On MIG-capable GPUs, commands such as nvidia-smi -q can display the GPU's current MIG Mode, including whether it is enabled or disabled. Therefore, nvidia-smi is the appropriate utility for verifying the GPU's MIG configuration.

🔴 Incorrect Options:

A. nvidia-mode
nvidia-mode is not a standard NVIDIA GPU management utility for checking MIG configuration. It cannot be used to verify whether MIG mode is enabled on a supported GPU.

B. nvidia-mig
nvidia-mig is not the standard NVIDIA utility for querying MIG mode. MIG configuration and status are managed and inspected using nvidia-smi commands.

C. nvidia-enable
nvidia-enable is not a standard NVIDIA GPU management command for checking MIG status. It does not provide the GPU configuration information needed to verify whether MIG mode is enabled.

🔧 Reference:
⇒ NVIDIA — MIG User Guide: Getting Started with MIG User Guide — Checking MIG Mode: documents using nvidia-smi to query and configure MIG mode on supported NVIDIA GPUs.

Explanation:

This question tests the purpose of NCCL (NVIDIA Collective Communications Library) in validating East-West fabric performance within an NVIDIA AI Factory. NCCL is the communication backbone for GPU-to-GPU collective operations, and its benchmarks are used to characterize both latency and bandwidth across the compute fabric.

✅ Correct Option B: To measure the latency between GPUs.
NCCL governs collective communication (all-reduce, all-gather, etc.) that requires tight synchronization across GPUs. The NCCL test suite measures how long data takes to traverse the GPU-to-GPU path, which directly reflects East-West fabric latency .

✅ Correct Option D: To measure bandwidth between GPUs.
NCCL benchmarks report bus bandwidth and algorithm bandwidth for large message sizes, revealing the effective throughput of the compute fabric connecting GPUs across nodes . This identifies whether the East-West fabric delivers expected capacity.

❌ Incorrect options:

A. To measure the storage network performance.
NCCL validates GPU-to-GPU collective communication over the compute (East-West) fabric, not storage network performance. Storage validation uses separate tools and targets North-South traffic .

C. To measure the power consumption of GPUs.
NCCL is a communication library with no power measurement capability. Power consumption is monitored through nvidia-smi or DCGM, not NCCL benchmarks.

🔧 Reference:
→ NVIDIA Multi-Node NVLink Benchmarking Guide
Confirms NCCL tests measure bandwidth and latency of collective operations across multi-GPU configurations.

→ NVIDIA NVL72 AI Factory Networking Topologies
States NCCL handles load balancing across compute planes and validates GPU-to-GPU connectivity in the East-West fabric.

Explanation:

This question tests your knowledge of NVIDIA BlueField Data Processing Unit (DPU) firmware update lifecycles. Flashing or upgrading the DPU firmware writes the new image to non-volatile flash memory, but the changes require a system reset to initialize.

✅ Correct Option:

✅ D. The firmware is waiting on reboot to become active.
When updating DPU firmware using tools like mlxconfig or flint, the new image is written to the secondary flash memory bank. The firmware remains in a pending state and will not take effect until a warm software reset (mlxreset) or a full system reboot/power cycle is performed to load the new image.

❌ Incorrect options:

❌ A. The firmware is installed on the DPU.
Stating that the firmware is fully installed implies that the new image is already running and active. While the files are written to flash memory, the running operational state remains on the older version until the system reboots.

❌ B. The firmware is deleted from the DPU.
Upgrading firmware writes new operational image data to the DPU's non-volatile memory. It does not erase or delete firmware without replacing it, as doing so would leave the device unbootable.

❌ C. The firmware is copied to the DPU but not installed.
The firmware update utility writes and verifies the firmware directly into the hardware flash registers, which completes the installation phase. It is not merely staged or copied as an uninstalled file.

🔧 Reference:
→ NVIDIA BlueField DPU Software Documentation confirms that updated firmware resides in flash memory and requires a reboot or reset to activate.

Explanation

The question tests the foundational stages of a recommended DGX cluster deployment. Nodes must be organized according to their roles, and all management, InfiniBand, and Ethernet interfaces must be configured and validated before installing workload-management software or running workloads.

✔️ Correct Option

A. Group nodes by function during initial setup and assign them to relevant categories in the cluster management tool.
Grouping nodes by function—such as head, login, compute, storage, or management roles—allows the cluster management tool to apply the correct images, configurations, networks, and policies. NVIDIA deployment procedures use node categories and assigned roles so that nodes boot with the settings appropriate for their purpose and can be managed consistently across the cluster.

B. Configure networking by validating all interfaces on each node, ensuring proper InfiniBand and Ethernet connectivity prior to installing cluster software.
Network validation is essential before deploying cluster software. Administrators should confirm that the required Ethernet and InfiniBand interfaces are detected, correctly configured, reachable, and connected to the intended networks. This establishes reliable management, storage, and GPU-to-GPU communication before installing workload-management or distributed-computing components.

❌ Incorrect Options

C. Install Slurm on the head node and then configure the compute nodes’ default OS images.
This reverses the recommended dependency order. Compute-node roles, images, interfaces, and network settings should be defined and validated before workload-manager deployment. Slurm installation is a later stage, after the nodes and underlying cluster infrastructure are properly provisioned.

D. Complete application containerization, run distributed jobs, and skip validation of node health or storage availability.
Application deployment and distributed job execution should occur only after infrastructure validation. Skipping node-health, storage, and network checks can hide hardware or connectivity failures and lead to unreliable job execution, performance problems, or data-access errors.

🔧 Reference
→ NVIDIA DGX BasePOD Deployment Guide
— Shows assigning DGX nodes to categories and images, configuring the cluster, and verifying node status after deployment.

→ NVIDIA DGX BasePOD Network Configuration
— Confirms configuring and verifying InfiniBand and node interfaces before completing cluster deployment.

Explanation

This question tests the first troubleshooting step for suspected GPU hardware problems on a DGX A100. Degraded GPU performance together with loud fans usually points to a thermal or power problem. The first step is to gather diagnostic data, using nvidia-smi to see GPU status, temperatures, clocks, and throttling reasons, before stressing, restarting, or changing the system.

✅ B. Check the NVIDIA System Management Interface (nvidia-smi) for GPU status and temperatures.
nvidia-smi shows each GPU's status, temperature, power draw, clocks, and any throttling reasons. Loud fans and reduced performance often mean a GPU is running hot and being throttled, and this output confirms it and shows which GPU is affected. It is non-disruptive, runs while the system is online, and gives the evidence needed to decide on further steps, such as NVSM health checks.

❌ A. Run a deep learning workload to stress test the GPUs and check whether the issue persists.
A stress test raises heat and power load on GPUs that are already suspected to be faulty or running hot. That can worsen the fault or trigger shutdowns, and it does not identify the cause. Diagnostic data such as temperatures and status should be collected first. Stress testing is better used later, in a controlled way, after the problem is understood.

❌ C. Power drain then restart the DGX and check if the performance degradation resolves.
Power draining and restarting takes the system offline and interrupts running workloads. It may temporarily hide the symptoms without finding the cause, and a real hardware fault, such as a failing fan or a GPU overheating, would return. It also removes the live state and evidence that could help pinpoint the fault. A restart is not a diagnostic step.

❌ D. Increase the fan speed to maximum and check whether the performance improves.
The fans are already running loudly, which suggests the system is responding to a thermal condition. Forcing them to maximum only masks the symptom and does not identify why the GPUs are hot. It also cannot fix a failed fan, blocked airflow, or a faulty GPU. It adds noise and wear without providing diagnostic information.

Reference
🔧 Quick Start and Basic Operation – NVIDIA DGX A100 User Guide → confirms that NVIDIA provides diagnostic tools (NVSM) to determine system health and identify component issues, which is used after checking GPU status and temperatures.