Free NCP-AII Practice Test Questions 2026

Explanation:

This question evaluates acceptance criteria for a multi-node NCCL all-reduce burn-in before production release. It tests whether engineers correctly distinguish transient, non-reproducible anomalies from persistent failures that block cluster handover, focusing on aggregate bandwidth stability, absence of hard errors, and recovery behavior over a long soak.

✅ Correct Option A:
Proceed is correct because aggregate bandwidth stays within 5 % of the hardware reference, NCCL logs show zero errors or timeouts, single-GPU dips recover immediately and remain stable, and WARN messages cannot be reproduced. These observations meet standard NCCL burn-in pass criteria for releasing the cluster.

❌ Incorrect Option B:
Recommend proactive maintenance is incorrect. Requiring zero variance above 10 % on any single run, even when transient and non-reproducible, is overly rigid. Long burn-ins routinely show brief, self-clearing fluctuations without hardware faults; rejecting the cluster on this basis would produce unnecessary false failures.

❌ Incorrect Option C:
Approve for AI workload use but flag affected nodes is incorrect. Flagging nodes for exclusion solely because of isolated, recovered dips and unreproducible WARNs lacks supporting evidence of ongoing risk. Isolation is warranted only for persistent or reproducible degradation, not for events that self-resolved with clean system logs.

🔧 Reference:
→ NVIDIA NCCL User Guide – Performance and Tuning – Confirms that healthy NCCL collective results require meeting expected bandwidth with no persistent errors or timeouts.

→ NVIDIA DGX BasePOD Deployment Guide – NCCL Validation – Shows multi-node NCCL all-reduce testing as a standard cluster validation step before production use.

Explanation:

This question tests how to monitor a BlueField-3 BMC firmware update after Redfish returns a "Running" task state. The key concept is that the Redfish TaskService maintains authoritative status information via Task ID polling, which is the standard NVIDIA-recommended method for tracking asynchronous firmware operations.

✅ Correct Option:

B. Query the DPU BMC with the Task ID of the installation process.
NVIDIA documentation confirms that firmware updates return a Redfish task URI, and the recommended method to track progress is polling /redfish/v1/TaskService/Tasks/ for PercentComplete, TaskState, and TaskStatus fields. A task showing "Running" with no progress after 15 minutes should be queried directly to determine if it is genuinely stalled or simply taking longer than expected, since offline update flows can remain at 0% until completion.

❌ Incorrect Options:

A. Check /var/log/messages on the DPU operating system for update logs.
This checks DPU Arm OS logs, not BMC firmware update status. BMC firmware updates are managed by the BMC's Redfish TaskService, not the DPU OS logging system. NVIDIA troubleshooting documentation directs users to BMC logs and Redfish task tracking for firmware operations.

C. Power cycle the DPU immediately to force a rollback.
Forcing a power cycle during an active firmware update risks corrupting the BMC firmware image and leaving the device in an unrecoverable state. NVIDIA documentation recommends monitoring task status until completion or documented failure before taking corrective action.

D. Run bfrec --status on the DPU to view flash progress.
There is no NVIDIA-documented command named bfrec --status for tracking BMC firmware update progress. The official method is Redfish TaskService polling. This option appears to be a fabricated distractor.

🔧 Reference:
→ NVIDIA BlueField BMC Software – Tracking Image Transfer Status: Documents that after receiving a running task state, users should poll /redfish/v1/TaskService/Tasks/ to track firmware update progress, with PercentComplete reaching 100 upon completion.

→ NVIDIA BlueField BMC Troubleshooting – Extract BMC Logs: Shows the identical Redfish task polling pattern for diagnostic data collection, confirming this is the authoritative status-tracking mechanism.

Explanation:

A “GPU fell off bus” error can indicate a GPU, PCIe, power, thermal, or related hardware problem. DCGM diagnostics provide targeted health and hardware checks while the burn-in is running, helping correlate failures with GPU telemetry and isolate the affected component.

🟢 Correct Option:

B. Run DCGM diagnostics alongside burn-in to monitor GPU health metrics.
DCGM diagnostics are designed to detect GPU health and hardware-related problems through targeted tests and monitoring. Running them alongside a multi-day burn-in can help correlate “GPU fell off bus” events with errors, thermal conditions, PCIe behavior, power, or other GPU health indicators. This provides evidence for determining whether the intermittent failures originate from hardware rather than the workload itself.

🔴 Incorrect Options:

A. Enable HPL_USE_NVSHMEM for alternative memory sharing.
NVSHMEM changes how memory is accessed and communicated in supported workloads. It does not provide the hardware health diagnostics needed to isolate a GPU, PCIe, power, or thermal fault causing a GPU to fall off the bus.

C. Switch from BERT to GPT models for simpler computations.
Changing the model does not directly diagnose hardware faults. It may alter workload characteristics, but it does not provide the GPU health telemetry or targeted diagnostic testing needed to identify the source of an intermittent bus failure.

D. Reduce blocksize to 500MB to lower memory pressure.
Reducing block size may change memory utilization and workload behavior, but it does not isolate the underlying hardware fault. A GPU falling off the bus requires hardware and system-level diagnostics rather than simply reducing memory pressure.

🔧 Reference:
⇒ NVIDIA DCGM Diagnostics — Confirms that DCGM provides diagnostic tests for assessing GPU health and detecting hardware-related issues.

Explanation:

This question tests your understanding of performance optimization techniques in the high-performance LINPACK benchmark for AI (HPL-AI) on NVIDIA DGX platforms. It evaluates how efficient GPU communication protocols minimize latency and maximize throughput across multi-GPU setups.

✅ Correct Option:

Option B:
HPL-AI leverages CUDA-aware MPI and GPUDirect communication protocols to transfer data directly between GPU memories across nodes and sockets. By bypassing the host CPU and system memory during collective communication, it significantly reduces CPU overhead, lowers latency, and keeps GPUs saturated, yielding massive performance gains on DGX clusters.

❌ Incorrect options:

Option A:
Reducing the problem size ($N$) decreases the overall computational load, which lowers the absolute time to completion but does not increase execution efficiency or floating-point throughput ($TFLOPS$) of the system.

Option C:
GPU clock speeds are governed by strict physical power, thermal, and architectural limits. Software or firmware updates cannot double GPU clock frequencies on hardware like the NVIDIA A100.

Option D:
NVLink bandwidth is determined by physical hardware interconnects and board architecture. Firmware and driver updates optimize bandwidth usage, but they cannot physically double the maximum physical link capacity of the hardware.

🔧 Reference:
→ NVIDIA HPL-AI Benchmark Guide
Confirms how CUDA-aware MPI and optimized inter-GPU communication drive performance gains in HPL-AI workloads.

Explanation:

This question assesses the correct early-stage validation sequence for an NVIDIA DGX system before enterprise production release. It focuses on catching hardware issues (such as intermittent GPU failures) through systematic hardware-first checks rather than skipping to software setup or user access.

✅ Correct Option B:
This sequence follows NVIDIA’s recommended pre-production process. After physical completion and power-on, firmware and drivers are updated, then NVSM performs full health checks and stress diagnostics. Sensor logs and GPU accessibility are verified to confirm reliability and performance consistency under load before any production use.

❌ Incorrect Option A:
Installing OS images and inviting users for a monitored trial is incorrect. It skips hardware-level stress diagnostics and sensor verification, allowing intermittent GPU failures to surface only after users are already working on the system.

❌ Incorrect Option C:
Focusing on network topology, static IPs, DNS, and basic OS checks is incorrect. These steps address connectivity but omit critical NVSM health checks, stress tests, and GPU sensor validation needed to prevent avoidable hardware-related downtime.

❌ Incorrect Option D:
Powering on and immediately installing AI frameworks and CUDA, then deferring stress tests, is incorrect. Diagnostics must occur early; postponing them until routine operations leaves latent hardware issues undetected until production impact occurs.

🔧 Reference:
→ NVIDIA DGX H100/H200 User Guide – Quickstart and Basic Operation – Recommends running the NVSM pre-flight stress test and health checks before placing a system into production.

→ NVIDIA System Management (NVSM) User Guide – Documents NVSM health checks, stress-test commands, and sensor verification for DGX production readiness.

Explanation:

This question tests how to physically connect a 25 GbE SFP28 host channel adapter to a QSFP port that supports both 100 GbE and 25 GbE. The core concept is using a passive form-factor adapter to bridge the smaller SFP28 connector into the larger QSFP28 cage.

✅ Correct Option:

C. QSA Adapter.
The NVIDIA LinkX QSA (QSFP-to-SFP Adapter) is a mechanical adapter that inserts into a QSFP port and accepts a single-channel SFP28 transceiver or cable. It enables a 25 GbE SFP28 device to connect to a QSFP28 port while passing only single-lane data, exactly matching the requirement for linking a 25 GbE HCA to a QSFP port.

❌ Incorrect Options:

A. SFP Connectors.
An SFP connector is the physical interface on the HCA side, not a solution for connecting to a QSFP port. SFP28 is the single-lane form factor used for 25 GbE, but it cannot mate directly with a QSFP28 cage without an adapter.

B. SFP to 1G BASE-T (RJ45) adapter.
This adapter converts an SFP port to copper RJ45 for 1 GbE connections only. It does not support 25 GbE speeds and does not bridge SFP28 to QSFP28. It is designed for legacy 1 GbE deployments, not 25 GbE links.

🔧 Reference:
→ NVIDIA LinkX Cable Accessories – QSA Overview: States QSA enables a single-channel SFP device to be inserted into a larger QSFP/QSFP-DD port, supporting 25G NRZ versions.

→ NVIDIA QSA Product Specification: Confirms the adapter connects an SFP28 transceiver to a QSFP28 port, with the SFP high-speed channel connected to lane 1 of the QSFP connector.

Explanation:

The NVSM health report marks Verify installed GPUs as Unhealthy and specifically reports a missing GPU at PCI address 07:00.0. This indicates that the system does not detect an expected GPU at that PCIe address. The output does not indicate a driver or GPU tray upgrade failure.

🟢 Correct Option:

B. A GPU is missing on the DGX system.
The health check explicitly identifies the GPU verification check as Unhealthy and states that lspci cannot find the expected GPU at PCI address 07:00.0. Because lspci reports PCI devices detected by the system, a missing expected GPU indicates that the hardware is not being detected at that address. The administrator should investigate the GPU, PCIe connection, or related hardware.

🔴 Incorrect Options:

A. A GPU tray upgrade failed.
The output does not mention a GPU tray upgrade or an unsuccessful upgrade operation. It only reports that an expected GPU is missing from PCI address 07:00.0.

C. A GPU driver upgrade has failed.
There is no driver-related error in the health report. The specific failure is an lspci check showing that an expected GPU is missing, which points to hardware detection rather than a driver upgrade.

D. The system has passed the hardware health check successfully.
This is incorrect because the GPU verification check is explicitly marked Unhealthy. Although the Ethernet, InfiniBand, PCIe switches, and DIMMs are healthy, the overall hardware health assessment cannot be considered successful with a missing GPU.

🔧 Reference:
⇒ NVIDIA System Management (NVSM) – Health Monitoring — Documents NVSM capabilities for monitoring and checking the health of DGX system components.

Explanation:

The question tests the correct way to fully purge all NVIDIA GPU driver packages and related components on Ubuntu 22.04 so that residual files or dependencies do not cause conflicts during troubleshooting.

✅ Correct Option: D. sudo apt-get purge nvidia-* && sudo apt-get autoremove
This command uses the purge flag with a wildcard to remove every package matching nvidia-* along with their configuration files. The subsequent autoremove cleans up any leftover dependencies. Together they ensure a complete and clean removal of the driver stack.

❌ Incorrect Option: A. sudo ubuntu-drivers uninstall
This is not a standard or complete uninstallation method. The ubuntu-drivers tool is mainly designed for driver detection and installation; it does not reliably purge all NVIDIA packages, configuration files, and residual components required for thorough troubleshooting.

❌ Incorrect Option: B. sudo rm -rf /usr/lib/nvidia
Manually deleting directories bypasses the package manager, leaves the system in an inconsistent state, and fails to remove all related packages, kernel modules, and configuration files. This approach is incomplete and can introduce further problems.

❌ Incorrect Option: C. sudo apt-get remove nvidia-driver-550
This command targets only one specific driver package and uses remove instead of purge, so configuration files and other NVIDIA packages remain on the system. It does not achieve a complete cleanup of all driver components.

🔧 Reference:
→ NVIDIA Driver Installation Guide – Removing the Driver (Ubuntu)
Confirms the use of apt remove --autoremove --purge on NVIDIA packages for a clean uninstall on Ubuntu 22.04.

Explanation:

The question tests how to improve storage-to-GPU data movement when CPU overhead is limiting throughput. NVIDIA GPUDirect Storage (GDS) creates a direct data path between storage and GPU memory, reducing unnecessary CPU involvement and helping GPUs receive data faster during checkpoint and dataset loading.

🟢 Correct Option:

D. Implement GPUDirect Storage to enable direct data transfers.
GPUDirect Storage allows data to move directly between GPU memory and supported storage without requiring the CPU to intermediate the transfer. This reduces CPU overhead and memory-copy operations while improving storage-to-GPU bandwidth. For workloads such as AI training and checkpoint restoration, GDS can help reduce GPU idle time caused by slow data loading and improve overall input/output efficiency.

🔴 Incorrect Options:

A. Increase batch sizes to reduce the frequency of storage access.
Larger batch sizes can reduce the relative frequency of some data-loading operations, but they do not remove CPU involvement in storage-to-GPU transfers. They also increase memory requirements and may not address the underlying storage bottleneck.

B. Migrate datasets to SATA SSDs with RAID 0 for higher sequential read speeds.
RAID 0 across SATA SSDs can improve storage throughput, but it does not provide direct storage-to-GPU transfers. CPU involvement and data-copy overhead can remain significant, so this does not directly address the stated CPU bottleneck.

C. Add more GPUs to the cluster to parallelize data loading tasks.
Adding GPUs increases compute capacity but does not inherently improve the storage-to-GPU data path. If existing GPUs are already waiting for data because of CPU and I/O overhead, adding more GPUs can increase contention for the same storage resources.

🔧 Reference:
⇒ NVIDIA GPUDirect Storage Documentation — NVIDIA documentation explains how GDS enables direct data movement between storage and GPU memory while reducing CPU involvement.

Explanation:

This question tests the correct Docker command syntax for restricting a container to specific NVIDIA GPUs by index. It verifies knowledge of the --gpus flag's device parameter and the required quoting format when enumerating multiple GPUs.

✅ Correct Option:

A. docker run --gpus '"device=0,2"' nvidia/cuda:12.1-base nvidia-smi
The --gpus flag with the device parameter is the standard method for exposing specific GPUs. Because the device list contains a comma, the value must be wrapped in single quotes followed by double quotes . This syntax correctly exposes only GPU indices 0 and 2 to the container.

❌ Incorrect options:

B. docker run -e NVIDIA_VISIBLE_DEVICES=0,2 nvidia/cuda:12.1-base nvidia-smi
While the NVIDIA_VISIBLE_DEVICES environment variable is valid, this command omits the --runtime=nvidia flag, which is typically required unless configured as the Docker default . Without it, the container may not properly receive GPU access.

C. docker run --gpus all nvidia/cuda:12.1-base nvidia-smi -id=0,2
This command exposes all GPUs to the container via --gpus all. The -id=0,2 flag is appended to nvidia-smi, not Docker, and nvidia-smi does not use this flag to restrict visible devices .

D. docker run --device /dev/nvidia0,/dev/nvidia2 nvidia/cuda:12.1-base nvidia-smi
The --device flag exposes raw device files without using the NVIDIA Container Toolkit. This bypasses the NVIDIA runtime's proper GPU enumeration and library mounting, often resulting in driver mismatches or permission issues .

🔧 Reference:
→ NVIDIA Container Toolkit – Specialized Configurations with Docker
Confirms the --gpus '"device=0,2"' syntax for exposing specific GPUs by index.

Explanation:

This question tests your understanding of data center efficiency measurement strategies for AI workloads. Standard infrastructure metrics fail to quantify the actual computational work delivered per unit of energy during complex machine learning jobs.

✅ Correct Option:

✅ C. Develop benchmarks tailored to specific workloads, such as MLPerf for AI applications, to better understand energy use in real-world scenarios.
Traditional facility metrics like PUE only measure power overhead ratios, not compute productivity. Workload-specific benchmarks like MLPerf compute-per-watt evaluate real-world performance directly against power consumed. Prioritizing workload benchmarks allows data center managers to accurately evaluate hardware, software stack optimizations, and throughput per joule during active AI training and inference.

❌ Incorrect options:

❌ A. Use Power Usage Effectiveness as the primary metric while supplementing it with additional measures of useful work done per unit of energy.
PUE measures the ratio of total facility power to IT equipment power. While useful for cooling infrastructure efficiency, relying on it as the primary metric fails because a facility can achieve an excellent PUE while running inefficient, low-throughput AI algorithms that waste massive amounts of energy.

❌ B. Use watts used as the primary measure of efficiency, as it accurately reflects the power input at any given time.
Watts measure instantaneous power consumption rather than efficiency. High power draw alone does not indicate inefficiency if the underlying GPUs complete high-density AI operations in significantly less time, delivering far higher work output per total unit of energy consumed.

❌ C. Focus on integrating kilowatt-hours into existing metrics to better reflect the actual energy used for productive work.
Kilowatt-hours measure raw energy consumption over time, but integrating raw energy units into facilities metrics still lacks context regarding computational output. Without tying energy units directly to standardized work tasks or training iterations, total energy usage remains an incomplete efficiency indicator.

🔧 Reference:
→ NVIDIA Data Center Efficiency and Sustainability confirms how workload-based performance metrics complement traditional facility power measurements.

Explanation

The question tests the NVIDIA Mellanox Firmware Tools command used to inspect link health and signal quality. mlxlink provides physical link counters, including BER information, while the -e option adds eye-opening diagnostics that help analyze signal integrity.

✔️ Correct Option

C. mlxlink -d -c -e
The mlxlink utility diagnoses device-link conditions. The -c option displays physical counters and BER information, while -e displays eye-opening data for additional signal-quality analysis. Together, these options provide BER-related link diagnostics and help identify physical-layer or signal-integrity problems on the selected device.

❌ Incorrect Options

A. mlxconfig -d q
mlxconfig is used to query or modify device configuration settings. The q option displays configuration information, but it does not collect link counters or measure BER.

B. ethtool -S
ethtool -S displays driver and network-interface statistics for an Ethernet device. Although some error counters may be available, it is not the NVIDIA-specific utility for collecting link BER and eye-opening diagnostics.

D. mstflint -d q full
mstflint is primarily used to query, burn, verify, and manage NVIDIA device firmware. The command can display firmware details, but it does not provide the physical link BER counters supplied by mlxlink.

🔧 Reference
→ NVIDIA mlxlink Utility
— Confirms that mlxlink can retrieve device information and BER counters.