NCP-AII NVIDIA Certified Professional AI Infrastructure Exam Free Dumps

dumps2025-09-06T10:09:09+08:00

NCP-AII NVIDIA Certified Professional AI Infrastructure exam is easy to take now. Learning NCP-AII exam information is a required step. There ar 65 questions in real NVIDIA NCP-AII exam, and you have 90 minutes to answer all the questions. The available language is English. The exam price is $400. NVIDIA Certified Professional AI Infrastructure certification is valid for two years. We provide the latest NVIDIA Certified Professional NCP-AII exam dumps, which can make sure you pass the test easily. Free NVIDIA Certified Professional AI Infrastructure NCP-AII exam dumps are available below.

Page 1 of 10

1. A GPU in your AI server consistently overheats during inference workloads. You’ve ruled out inadequate cooling and software bugs.

Running ‘nvidia-smi’ shows high power draw even when idle.

Which of the following hardware issues are the most likely causes?

Degraded thermal paste between the GPU die and the heatsink.

A failing voltage regulator module (VRM) on the GPU board, causing excessive power leakage.

Incorrectly seated GPU in the PCle slot, leading to poor power delivery.

A BIOS setting that is overvolting the GP

Insufficient system RA

2. You are monitoring a server with 8 GPUs used for deep learning training. You observe that one of the GPUs reports a significantly lower utilization rate compared to the others, even though the workload is designed to distribute evenly. ‘nvidia-smi’ reports a persistent "XID 13" error for that GPU.

What is the most likely cause?

A driver bug causing incorrect workload distribution.

Insufficient system memory preventing data transfer to that GP

A hardware fault within the GPU, such as a memory error or core failure.

An incorrect CUDA version installed.

The GPU’s compute mode is set to ‘Exclusive Process’.

3. You notice that one of the fans in your GPU server is running at a significantly higher RPM than the others, even under minimal load. ipmitool sensor’ output shows a normal temperature for that GPU.

What could be the potential causes?

The fan’s PWM control signal is malfunctioning, causing it to run at full speed.

The fan bearing is wearing out, causing increased friction and requiring higher RPM to maintain airflow.

The fan is attempting to compensate for restricted airflow due to dust buildup.

The server’s BMC (Baseboard Management Controller) has a faulty temperature sensor reading, causing it to overcompensate.

A network connectivity issue is causing higher CPU utilization, leading to increased system-wide heat.

4. After upgrading the network card drivers on your A1 inference server, you experience intermittent network connectivity issues, including packet loss and high latency. You’ve verified that the physical connections are secure.

Which of the following steps would be most effective in troubleshooting this issue?

Roll back the network card drivers to the previous version.

Check the system logs for error messages related to the network card or driver.

Run network diagnostic tools like ‘ping’, ‘traceroute’, and ‘iperf3’ to assess the network performance.

Reinstall the operating system.

Update the server’s BIO

5. Your deep learning training job that utilizes NCCL (NVIDIA Collective Communications Library) for multi-GPU communication is failing with "NCCL internal error, unhandled system error" after a recent CUDA update. The error occurs during the ‘all reduce’ operation.

What is the most likely root cause and how would you address it?

Incompatible NCCL version with the new CUDA version. Update NCCL to a version compatible with the installed CUDA version.

Insufficient shared memory allocated to the CUDA context. Increase the shared memory limit using ‘cudaDeviceSetLimit(cudaLimitSharedMemory, new_limity.

Firewall rules blocking inter-GPU communication. Configure the firewall to allow communication on the NCCL-defined ports (typically 8000-8010).

Faulty network cables used for inter-node communication (if the training job spans multiple servers). Replace the network cables with certified high-speed cables.

GPU Direct RDMA is not properly configured. Check ‘dmesg’ for errors and ensure RDMA is enabled.

6. You are deploying a new A1 inference service using Triton Inference Server on a multi-GPU system. After deploying the models, you observe that only one GPU is being utilized, even though the models are configured to use multiple GPUs.

What could be the possible causes for this?

The model configuration file does not specify the ‘instance_group’ parameter correctly to utilize multiple GPUs.

The Triton Inference Server is not configured to enable CUDA Multi-Process Service (MPS).

Insufficient CPU cores are available for the Triton Inference Server, limiting its ability to spawn multiple inference processes.

The models are not optimized for multi-GPU inference, resulting in a single GPU bottleneck.

The GPUs are not of the same type and Triton cannot properly schedule across them.

7. An A1 server exhibits frequent kernel panics under heavy GPU load. ‘dmesg’ reveals the following error: ‘NVRM: Xid (PCl:0000:3B:00): 79, pid=..., name=..., GPU has fallen off the bus.’

Which of the following is the least likely cause of this issue?

Insufficient power supply to the GPIJ, causing it to become unstable under load.

A loose or damaged PCle riser cable connecting the GPU to the motherboard.

A driver bug in the NVIDIA drivers, leading to GPU instability.

Overclocking the GPU beyond its stable limits.

A faulty CP

8. You are using GPU Direct RDMA to enable fast data transfer between GPUs across multiple servers. You are experiencing performance degradation and suspect RDMA is not working correctly.

How can you verify that GPU Direct RDMA is properly enabled and functioning?

Check the output of ‘nvidia-smi topo -m’ to ensure that the GPUs are connected via NVLink and have RDMA enabled.

Examine the ‘cimesg’ output for any errors related to RDMA or InfiniBand drivers.

Use the ‘ibstat command to verify that the InfiniBand interfaces are active and connected.

Run a bandwidth benchmark using a tool like or to measure the RDMA throughput.

Ping the other servers to ensure network connectivity.

9. You’re debugging performance issues in a distributed training job. ‘nvidia-smi’ shows consistently high GPU utilization across all nodes, but the training speed isn’t increasing linearly with the number of GPUs. Network bandwidth is sufficient.

What is the most likely bottleneck?

Inefficient data loading and preprocessing pipeline, causing GPUs to wait for data.

NCCL is not configured optimally for the network topology, leading to high communication overhead.

The learning rate is not adjusted appropriately for the increased batch size across multiple GPUs.

The global batch size has exceeded the optimal point for the model, reducing per-sample accuracy and slowing convergence.

CUDA Graphs is not being utilized.

10. You have a server equipped with multiple NVIDIA GPUs connected via NVLink. You want to monitor the NVLink bandwidth utilization in real-time.

Which tool or method is the most appropriate and accurate for this?

Using ‘nvidia-smi’ with the ‘―display=nvlink’ option.

Parsing the output of *nvprof during a representative workload.

Utilizing DCGM (Data Center GPU Manager) with its NVLink monitoring capabilities.

Monitoring network interface traffic using ‘iftop’ or ‘tcpdump’ .

Using ‘gpustat’ .

Page 2 of 10

11. Your A1 inference server utilizes Triton Inference Server and experiences intermittent latency spikes. Profiling reveals that the GPU is frequently stalling due to memory allocation issues.

Which strategy or tool would be least effective in mitigating these memory allocation stalls?

Using CIJDA memory pools to pre-allocate memory and reduce allocation overhead during inference requests.

Enabling CUDA graph capture to reduce kernel launch overhead.

Reducing the model’s memory footprint by using quantization or pruning techniques.

Increasing the GPU’s TCC (Tesla Compute Cluster) mode priority.

Optimize the model using TensorR

12. You are tasked with troubleshooting a performance bottleneck in a multi-node, multi-GPU deep learning training job utilizing Horovod.

The training loss is decreasing, but the overall training time is significantly longer than expected.

Which of the following monitoring approaches would provide the most insight into the cause of the bottleneck?

Using ‘nvidia-smi’ on each node to monitor GPU utilization and memory usage.

Enabling Horovod’s timeline and profiling features to visualize the communication patterns and identify synchronization bottlenecks.

Monitoring network bandwidth utilization on each node using ‘iftop’ or ‘iperf3’

Analyzing the training loss curve to identify potential issues with the model architecture or hyperparameters.

Using Shtop’ to monitor CPIJ utilization on each node.

13. You’re troubleshooting a DGX-I server exhibiting performance degradation during a large-scale distributed training job. ‘nvidia-smü shows all GPUs are detected, but one GPU consistently reports significantly lower utilization than the others. Attempts to reschedule orkloads to that GPU frequently result in CUDA errors.

Which of the following is the MOST likely cause and the BEST initial roubleshooting step?

A driver issue affecting only one GPU; reinstall NVIDIA drivers completely.

A software bug in the training script utilizing that specific GPU’s resources inefficiently; debug the training script.

A hardware fault with the GPU, potentially thermal throttling or memory issues; run ‘nvidia-smi -i -q’ to check temperatures, power limits, and error counts.

Insufficient cooling in the server rack; verify adequate airflow and cooling capacity for the rack.

Power supply unit (PSU) overload, causing reduced power delivery to that GPU; monitor PSU load and check PSU specifications.

14. An A1 inferencing server, using NVIDIA Triton Inference Server, experiences intermittent crashes under peak load. The logs reveal CUDA out-of-memory errors (00M) despite sufficient system RAM. You suspect a GPU memory leak within one of the models.

Which strategy BEST addresses this issue?

Increase the system RAM to accommodate the growing memory footprint.

Implement CUDA memory pooling within the Triton Inference Server configuration to reuse memory allocations efficiently.

Reduce the batch size and concurrency of the offending model in the Triton configuration.

Upgrade the GPUs to models with larger memory capacity.

Disable other models running on the same GPU to free up memory.

15. You are managing a cluster of GPU servers for deep learning. You observe that one server consistently exhibits high GPU temperature during training, causing thermal throttling and reduced performance. You’ve already ensured adequate airflow.

Which of the following actions would be MOST effective in addressing this issue?

Reduce the ambient temperature of the data center.

Lower the GPU power limit using ‘nvidia-smi ―power-limit*.

Update the NVIDIA drivers to the latest version.

Re-seat the GPU in its PCle slot to ensure proper contact and heat dissipation.

Increase the fan speed of the GPU cooler using ‘nvidia-smi --fan’.

16. A DGX A100 server with dual power supplies reports a critical power event in the BMC logs. One PSU shows a ‘degraded’ status, while the other appears normal.

What immediate actions should you take to ensure continued operation and prevent data loss?

Immediately shut down the server gracefully to prevent further damage to the faulty PSI

Hot-swap the degraded PSU with a replacement unit.

Monitor the remaining PSU’s load and temperature closely; if stable, continue operation until a scheduled maintenance window.

Reduce the GPU power limit using ‘nvidia-smi’ to decrease the overall power consumption of the server.

Migrate all workloads to other servers in the cluster to minimize the impact of a potential complete PSU failure.

17. You are running a distributed training job on a multi-GPU server. After several hours, the job fails with a NCCL (NVIDIA Collective Communications Library) error. The error message indicates a failure in inter-GPU communication. ‘nvidia-smi’ shows all GPUs are healthy.

What is the MOST probable cause of this issue?

A bug in the NCCL library itself; downgrade to a previous version of NCC

Incorrect NCCL configuration, such as an invalid network interface or incorrect device affinity settings.

Insufficient inter-GPU bandwidth; reduce the batch size to decrease communication overhead.

A faulty network cable connecting the server to the rest of the cluster.

Driver incompatibility issue between NCCL and the installed NVIDIA driver version.

18. Your AI infrastructure includes several NVIDIAAI 00 GPUs. You notice that the GPU memory bandwidth reported by ‘nvidia-smi’ is significantly lower than the theoretical maximum for all GPUs. System RAM is plentiful and not being heavily utilized.

What are TWO potential bottlenecks that could be causing this performance issue?

Insufficient CPU cores assigned to the training process.

Inefficient data loading from storage to GPU memory.

The GPUs are connected via PCle Gen3 instead of PCle Gen4.

The CPU is using older DDR4 memory with low bandwidth

The NVIDIA drivers are not configured to enable peer-to-peer memory access between GPUs.

19. You suspect a faulty NVIDIA ConnectX-6 network adapter in a server used for RDMA-based distributed training.

Which commands or tools can you use to diagnose potential issues with the adapter’s hardware and connectivity?

Ispci -v to verify the adapter is detected and its resources are allocated correctly.

ibstat to check the adapter’s status, link speed, and active ports.

ethtool to examine the adapter’s Ethernet settings and statistics.

ping to test basic network connectivity.

nvsmimonitord to monitor GPU metrics and detect anomalies.

20. An AI server with 8 GPUs is experiencing random system crashes under heavy load. The system logs indicate potential memory errors, but standard memory tests (memtest86+) pass without any failures. The GPUs are passively cooled.

What are the THREE most likely root causes of these crashes?

Incompatible NVIDIA driver version with the installed Linux kernel.

GPIJ memory errors that are not detectable by standard CPU-based memory tests.

Insufficient airflow within the server, leading to overheating of the GPUs and VRMs.

A faulty power supply unit (PSU) that is unable to provide stable power under peak load.

Network congestion causing intermittent data corruption during distributed training.

Page 3 of 10

21. Consider a scenario where you are running a CUDA application on an NVIDIA GPU. The application compiles successfully but crashes during runtime with a *CUDA ERROR ILLEGAL ADDRESS* error. You’ve carefully reviewed your code and can’t find any obvious out- of-bounds memory accesses.

What advanced debugging techniques could help you pinpoint the source of this error?

Use ‘cuda-memcheck’ to detect memory access errors at runtime.

Employ the CUDA Debugger (cuda-gdb) to step through the code and inspect variable values and memory contents.

Utilize NVIDIA Nsight Systems to profile the application and identify memory allocation patterns.

Enable ECC (Error Correction Code) memory on the GPU to detect and correct memory errors.

Reduce the block size used in CUDA kernels to decrease the likelihood of shared memory conflicts.

22. A critical AI model training job consistently fails on a specific GPU server in your cluster after running for approximately 24 hours.

Monitoring data shows a sudden drop in GPU power consumption followed by a system reboot. All other GPUs on the server appear normal. The server has redundant PSUs.

What is the MOST likely cause?

A software bug in the A1 model causing a kernel panic specifically triggered after 24 hours of execution.

Thermal runaway on the GPU due to a failing thermal interface material (TIM) between the GPU die and the heatsink.

A transient power supply issue affecting only one of the redundant PSUs, triggering a system-wide protection mechanism.

ECC memory errors accumulating over time, eventually leading to a non-recoverable system fault.

A driver crash, causing the GPU to reset and the system to reboot.

23. You’re optimizing a deep learning model for deployment on NVIDIA Tensor Cores. The model uses a mix of FP32 and FP16 precision. During profiling with NVIDIA Nsight Systems, you observe that the Tensor Cores are underutilized.

Which of the following strategies would MOST effectively improve Tensor Core utilization?

Increase the batch size to fully utilize the available GPU memory.

Ensure that all matrix multiplications are performed using FP16 precision.

Pad the input tensors to dimensions that are multiples of 8 for optimal Tensor Core alignment.

Enable CUDA graph capture to reduce kernel launch overhead.

Decrease the learning rate to improve training stability and reduce the need for gradient clipping.

24. You are tasked with diagnosing performance issues on a GPU server running a large-scale HPC simulation. The simulation utilizes multiple GPUs and InfiniBand for inter-GPU communication. You suspect that RDMA (Remote Direct Memory Access) is not functioning correctly.

How would you comprehensively test and verify the proper operation of RDMA between the GPUs?

Use ‘ping’ to verify basic network connectivity between the server’s InfiniBand interfaces.

Employ and from the ‘perftest’ suite to measure RDMA bandwidth and latency between GPUs.

Run ‘nvidia-smi topo -m’ to check the GPU interconnect topology and verify that NVLink or PCle is being used for communication.

Utilize NCCL’s internal diagnostic tools to verify proper inter-GPU communication within the simulation.

Monitor CPU utilization during the simulation; high CPU usage suggests that RDMA is not offloading communication effectively.

25. A server with eight NVIDIAAIOO GPUs experiences frequent CUDA errors during large model training. ‘nvidia-smi’ reports seemingly normal temperatures for all GPUs. However, upon closer inspection using IPMI, the inlet temperature for GPUs 3 and 4 is significantly higher than others.

What is the MOST likely cause and the immediate action to take?

A driver issue is causing incorrect temperature reporting; reinstall the NVIDIA driver.

The temperature sensors on GPUs 3 and 4 are faulty; replace the GPUs immediately.

There is a localized airflow problem affecting GPUs 3 and 4; check fan speeds and airflow obstructions.

The power supply is failing to provide sufficient power to GPUs 3 and 4; replace the power supply.

A software bug in the CUDA toolkit is causing the errors; downgrade to an earlier version.

26. You are replacing a faulty NVIDIA Tesla V 100 GPU in a server. After physically installing the new GPU, the system fails to recognize it. You’ve verified the power connections and seating of the card.

Which of the following steps should you take next to troubleshoot the issue?

Immediately RMA the new GPU as it is likely defective.

Update the system BIOS and BMC firmware to the latest versions.

Reinstall the operating system to ensure proper driver installation.

Check if the new GPU requires a different driver version than the currently installed one and update if needed.

Disable and re-enable the GPU slot in the system BIO

27. A data scientist reports that training performance on a DGX A100 server has significantly degraded over the past week. ‘nvidia-smi’ shows all GPUs functioning, but ‘nvprof’ reveals substantially increased ‘cudaMemcpy’ times.

What is the MOST likely bottleneck?

The CPU is heavily loaded, causing contention for system memory bandwidth.

The PCle bus is saturated, limiting data transfer speeds between the CPU and GPUs.

The NVLink connections between GPUs are failing, forcing data transfers through PCle.

The GPUs are overheating, causing thermal throttling and slower memory transfers.

The storage system is slow, delaying data loading and preprocessing.

28. You are tasked with replacing a redundant power supply unit (PSU) in a GPU server. The server has two 2000W PSUs. One PSU has failed, but the server is still running.

Which of the following actions is the safest and most efficient way to replace the faulty PSU?

Immediately power down the server and replace the faulty PSI

Hot-swap the faulty PSU with a new one while the server is running.

Wait for a scheduled maintenance window to power down the server and replace the PS

Replace the faulty PSU, then reboot the server to ensure the new PSU is working.

Document the failure and wait until the remaining PSU fails before taking action.

29. After replacing a GPU in a multi-GPU server, you notice that the new GPU is consistently running at a lower clock speed than the other GPUs, even under load. *nvidia-smi’ shows the ‘Pwr’ state as ‘P8’ for the new GPU, while the others are at ‘PO’.

What is the MOST probable cause?

The new GPU is a lower-performance model than the other GPUs.

The driver is not properly recognizing the new GPU’s capabilities; reinstall the driver.

The new GPU is not receiving sufficient power; check the power connections and PSU capacity.

The new GPU is overheating and throttling performance.

The new GPU requires a firmware update that hasn’t been applied.

30. You’re deploying a new cluster with multiple NVIDIAAIOO GPUs per node. You want to ensure optimal inter-GPU communication performance using NVLink.

Which of the following configurations are critical for achieving maximum NVLink bandwidth?

All GPUs within a node must be the same model and have identical firmware versions.

The motherboard must support PCle Gen5 to maximize NVLink bandwidth.

GPUs should be physically installed in slots that maximize direct NVLink connections based on the server’s architecture.

The NVIDIA driver must be configured to enable NVLink; it is disabled by default.

The server must use a specific CPU model to leverage NVLink capabilities.

Page 4 of 10

31. A user reports that their deep learning training job is crashing with a ‘CUDA out of memory’ error, even though ‘nvidia-smi’ shows plenty of free memory on the GPU. The job uses TensorFlow.

What are the TWO most likely causes?

The TensorFlow version is incompatible with the installed NVIDIA driver.

TensorFlow is allocating memory on the CPU instead of the GP

TensorFlow is fragmenting GPU memory, making it difficult to allocate contiguous blocks.

The CUDA VISIBLE DEVICES environment variable is not set correctly.

The system’s swap space is full, preventing memory from being allocated.

32. You suspect a power supply issue is causing intermittent GPU failures in a server with four NVIDIAAIOO GPUs. The server is rated for a peak power consumption of 3000W. You have a power meter available.

Which of the following methods provides the most accurate assessment of the server’s power consumption under full GPU load?

Run ‘nvidia-smi’ and sum the reported power consumption for each GPI

Use the power meter to measure the server’s power consumption at idle and multiply by four.

Use the power meter to measure the server’s power consumption while running a synthetic benchmark that fully utilizes all GPIJs simultaneously.

Check the server’s BIOS for power consumption readings.

Add the maximum power rating of each GPU to the CPU’s TDP (Thermal Design Power).

33. A distributed training job using multiple nodes, each with eight NVIDIA GPUs, experiences significant performance degradation. You notice that the network bandwidth between nodes is consistently near its maximum capacity. However, ‘nvidia-smi’ shows low GPU utilization on some nodes.

What is the MOST likely cause?

The GPUs are overheating, causing thermal throttling.

Data is not being distributed evenly across the nodes; some nodes are waiting for data from others.

The NVIDIA drivers are outdated, causing communication bottlenecks.

The network interface cards (NICs) are faulty, causing packet loss and retransmissions.

The CPU is heavily loaded, causing contention for network resources.

34. After replacing a faulty NVIDIA GPU, the system boots, and ‘nvidia-smi’ detects the new card. However, when you run a CUDA program, it fails with the error "‘no CUDA-capable device is detected’". You’ve confirmed the correct drivers are installed and the GPU is properly seated.

What’s the most probable cause of this issue?

The new GPU is incompatible with the existing system BIO

The CUDA toolkit is not properly configured to use the new GP

The ‘LD LIBRARY PATH* environment variable is not set correctly.

The user running the CUDA program does not have the necessary permissions to access the GP

The GPIJ is not properly initialized by the system due to a missing or incorrect ACPI configuration.

35. You are managing a server farm of GPU servers used for A1 model training. You observe frequent GPU failures across different servers.

Analysis reveals that the failures often occur during periods of peak ambient temperature in the data center. You can’t immediately improve the data center cooling.

What are TWO proactive measures you can implement to mitigate these failures without significantly impacting training performance?

Reduce the GPU power limit using ‘nvidia-smi’ to decrease heat generation.

Increase the fan speeds of the GPU coolers to improve heat dissipation.

Implement a more aggressive GPU frequency scaling profile to throttle performance during peak temperatures.

Schedule training jobs to run during off-peak hours when ambient temperatures are lower.

Replace all existing GPUs with water-cooled models.

36. You’ve replaced a faulty NVIDIA Quadro RTX 8000 GPU with an identical model in a workstation. The system boots, and ‘nvidia-smi’ recognizes the new GPU. However, when rendering complex 3D scenes in Maya, you observe significantly lower performance compared to before the replacement. Profiling with the NVIDIA Nsight Graphics debugger shows that the GPU is only utilizing a small fraction of its available memory bandwidth.

What are the TWO most likely contributing factors?

The new GPU’s PCle link speed is operating at a lower generation (e.g., Gen3 instead of Gen4).

The NVIDIA OptiX denoiser is not properly configured or enabled.

The workstation’s power plan is set to ‘Power Saver,’ limiting GPU performance.

The Maya scene file contains corrupted or inefficient geometry.

The newly installed GPU’s VBIOS has not been properly flashed, causing an incompatibility issue.

37. You’re optimizing an AMD EPYC server with 4 NVIDIAAIOO GPUs for a large language model training workload. You observe that the GPUs are consistently underutilized (50-60% utilization) while the CPUs are nearly maxed out.

Which of the following is the MOST likely bottleneck?

Insufficient CPU cores to prepare and feed data to the GPUs.

The PCle interconnect between the CPUs and GPUs is saturated.

The system RAM is too small, causing excessive swapping.

The storage system (SSD/NVMe) is too slow, leading to data starvation.

The NCCL (NVIDIA Collective Communications Library) is not properly configured for inter-GPU communication.

38. Which of the following techniques are effective for improving inter-GPU communication performance in a multi-GPU Intel Xeon server used for distributed deep learning training with NCCL?

Enabling PCle peer-to-peer transfers between GPUs.

Utilizing InfiniBand or RoCE interconnects if available.

Increasing the system RAM size to minimize data transfer to disk.

Configuring NCCL to use the correct network interface and transport protocol (e.g., 1B, Socket).

Disabling CPU frequency scaling to maintain consistent performance.

39. You’re profiling the performance of a PyTorch model running on an AMD server with multiple NVIDIA GPUs. You notice significant overhead in the data loading pipeline.

Which of the following strategies can help optimize data loading and improve GPU utilization? Select all that apply.

Using the ‘torch.utils.data.DataLoader’ with multiple worker processes.

Loading the entire dataset into RAM before training.

Implementing asynchronous data prefetching using ‘torch .Generator’.

IJsing a faster storage system (e.g., NVMe SSD instead of HDD).

Reducing the batch size to decrease the amount of data loaded per iteration.

40. You have an Intel Xeon Gold server with 2 NVIDIA Tesla VI 00 GPUs. After deploying your A1 application, you observe that one GPU is consistently running at a significantly higher temperature than the other

What could be a plausible reason for this behavior?

One GPU is defective and drawing excessive power.

The server’s airflow is inadequate, causing poor cooling for one of the GPUs.

The workload is not evenly distributed between the GPUs, causing one GPU to be more heavily utilized.

One GPU’s driver version is outdated, leading to inefficient power management.

The ambient temperature in the server room is higher on one side of the rack.

Page 5 of 10

41. You are tasked with optimizing an Intel Xeon scalable processor-based server running a TensorFlow model with multiple NVIDIA GPUs.

You observe that the CPU utilization is low, but the GPU utilization is also not optimal. The profiler shows significant time spent in ‘tf.data’ operations.

Which of the following actions would MOST likely improve performance?

Increase the number of threads used for CPU-bound operations in TensorFlow using ‘tf.config.threading.set_intra_op_parallelism_threads()’.

Enable XLA (Accelerated Linear Algebra) compilation in TensorFlow.

Use ‘tf.data.AUTOTIJNE to allow TensorFlow to dynamically optimize the data pipeline.

Reduce the global batch size to improve memory utilization.

Upgrade the server’s network adapter to a faster interface, such as 100Gb

42. Consider an AMD EPYC-based server with 8 NVIDIAAIOO GPUs connected via PCle Gen4. You’re running a distributed training job using Horovod. You’ve noticed that communication between GPUs is a bottleneck.

Which of the following NCCL configuration options would be MOST beneficial in this scenario? (Assume all options are syntactically correct for NCCL).

NCCL SOCKET IF-NAME=eth0

NCCL IB DISABLE-1

NCCL P2P DISABLE-0

NCCL IB HCA=mlx5 0

NCCL NET PLUGIN=none

43. You’re optimizing an Intel Xeon server with 4 NVIDIA GPUs for inference serving using Triton Inference Server. You’ve deployed multiple models concurrently. You observe that the overall throughput is lower than expected, and the GPU utilization is not consistently high.

What are potential bottlenecks and optimization strategies? (Select all that apply)

Model loading and unloading overhead. Use model ensemble or dynamic batching to reduce frequency.

Insufficient CPU cores to handle the model loading and preprocessing requests. Increase the number of Triton instance groups for CPU-based models.

The models are memory-bound. Reduce the model precision (e.g., FP32 to FP16 or INT8).

The GPUs are underutilized due to small batch sizes. Implement dynamic batching to increase batch sizes.

Insufficient PCle bandwidth between CPU and GPIJs. Reconfigure PCle lanes to improve bandwidth allocation to each GPI

44. Which of the following are key considerations when choosing between CPU pinning and NUMA (Non-Uniform Memory Access) awareness for a distributed training job on a multi-socket AMD EPYC server with multiple GPUs?

CPU pinning ensures that each process/thread runs on a specific CPU core, reducing context switching overhead. NUMA awareness ensures that the CPU cores and memory used by a process are located within the same NUMA node, minimizing memory access latency.

CPU pinning is generally more important than NIJMA awareness because it directly impacts CPU utilization.

NUMA awareness is generally more important than CPU pinning because it directly impacts memory bandwidth.

Both CPU pinning and NUMA awareness are critical for optimizing performance. They should be used in conjunction to achieve optimal performance.

Neither CPU pinning nor NUMA awareness are relevant for GPIJ-accelerated workloads, as the GPUs handle all the computation.

45. You are troubleshooting a performance issue on an Intel Xeon server with NVIDIAAI 00 GPUs. Your application involves frequent data transfers between CPU memory and GPU memory. You suspect that the PCle bus is a bottleneck.

How can you verify and mitigate this bottleneck?

Use ‘nvidia-smi’ to monitor the PCle bandwidth utilization of the GPUs. If it’s consistently high (near the theoretical limit), the PCle bus is likely a bottleneck. Mitigate by reducing the frequency of CPU-GPU data transfers, using pinned (page-locked) memory, and ensuring that the GPUs are connected to PCle slots with sufficient bandwidth.

Check the CPU utilization. If it’s low, the PCle bus is likely the bottleneck. Mitigate by increasing the number of CPU cores assigned to the data transfer tasks.

Examine the system logs for PCle errors. If there are many errors, the PCle bus is likely unstable. Mitigate by reseating the GPUs and checking the power supply.

Monitor the GPU temperature. If it’s high, the PCle bus is likely overheating. Mitigate by improving the server’s cooling.

Use ‘nvprof to profile the application and identify the exact lines of code that are causing the high PCle traffic. Optimize those sections of code to reduce data transfers.

46. You have a large dataset stored on a network file system (NFS) and are training a deep learning model on an AMD EPYC server with NVIDIA GPUs. Data loading is very slow.

What steps can you take to improve the data loading performance in this scenario? Select all that apply.

Increase the number of NFS client threads on the AMD EPYC server.

Use a local SSD or NVMe drive to cache frequently accessed data.

Mount the NFS share with the ‘nolock’ option.

Switch to a parallel file system like Lustre or BeeGF

Reduce the batch size to decrease the amount of data loaded per iteration.

47. You’re optimizing an Intel Xeon server with 4 NVIDIAAIOO GPUs for a computer vision application that uses CODA. You notice that the GPU utilization is fluctuating significantly, and performance is inconsistent. Using ‘nvprof, you identify that there are frequent stalls in the CUDA kernels due to thread divergence.

What are possible causes and solutions?

The input data is not properly aligned in memory. Ensure that data is aligned to 128-byte boundaries using aligned memory allocation techniques.

The CUDA code contains conditional branches that lead to different execution paths for different threads within the same warp. Rewrite the CUDA code to minimize branching and favor uniform execution paths within warps.

The GPUs are overheating, causing thermal throttling. Improve the server’s cooling.

The CUDA compiler is generating suboptimal code. Try using different compiler optimization flags (e.g., ‘-O3’ or ‘-ftz=true’).

The CUDA driver version is incompatible with the CUDA toolkit version. Update the CUDA driver to a compatible version.

48. You are running a large-scale distributed training job on a cluster of AMD EPYC servers, each equipped with multiple NVIDIAA100 GPUs. You are using Slurm for job scheduling. The training process often fails with NCCL errors related to network connectivity.

What steps can you take to improve the reliability of the network communication for NCCL in this environment? Choose the MOST appropriate answers.

Ensure that the InfiniBand or RoCE network is properly configured and that all servers can communicate with each other over the network. Verify the network interface names and IP addresses in the NCCL configuration.

Use the Slurm ‘srun’ command with the ‘―mpi=pmi2 option to launch the training job. This ensures that Slurm properly initializes the MPl environment and sets the NCCL environment variables.

Increase the ‘NCCL CONNECT TIMEOUT and *NCCL TIMEOUT environment variables to allow for longer network delays.

Disable the firewall on all servers to allow unrestricted network communication.

Decrease the batch size to reduce the amount of data transferred over the network.

49. You are tasked with optimizing storage performance for a deep learning training job on an NVIDIA DGX server. The training data consists of millions of small image files.

Which of the following storage optimization techniques would be MOST effective in reducing I/O bottlenecks?

Implementing RAID 0 across all storage devices.

Using a distributed file system with data striping across multiple storage nodes.

Enabling data compression on the storage volume.

Increasing the block size of the file system to the maximum supported value.

Implementing a tiered storage system with NVMe drives for frequently accessed data and HDDs for less frequently accessed data.

50. A data scientist reports slow data loading times when training a large language model. The data is stored in a Ceph cluster. You suspect the client-side caching is not properly configured.

Which Ceph configuration parameter(s) should you investigate and potentially adjust to improve data loading performance? Select all that apply.

client cache size

client quota

mds cache size

fuse_client_max_background

Page 6 of 10

51. You’re monitoring the storage I/O for an AI training workload and observe high disk utilization but relatively low CPU utilization.

Which of the following actions is LEAST likely to improve the performance of the training job?

Switching from HDDs to NVMe SSDs.

Implementing data prefetching to load data into memory before it’s needed.

Increasing the batch size of the training job.

Adding more RAM to the system.

Reducing the number of parallel data loading threads.

52. Which of the following is the MOST important reason for using a dedicated storage network (e.g., InfiniBand or RoCE) for AI/ML workloads compared to using the existing Ethernet network?

Improved security due to network isolation.

Lower latency and higher bandwidth for data transfer.

Simplified network management and configuration.

Reduced cost compared to upgrading the existing Ethernet infrastructure.

Automatic Quality of Service (QOS) prioritization for AI/ML traffic.

53. Your AI training pipeline involves a pre-processing step that reads data from a large HDF5 file. You notice significant delays during this step. You suspect the HDF5 file structure might be contributing to the slow read times.

What optimization technique is MOST likely to improve read performance from this HDF5 file?

Converting the HDF5 file to a CSV file.

Storing the HDF5 file on a network file system like NF

Reorganizing the HDF5 file to improve data contiguity and chunking.

Compressing the HDF5 file using gzip.

Encrypting the HDF5 file for enhanced security.

54. You are running a distributed training job across multiple nodes, using a shared file system for storing training dat

a. You observe that some nodes are consistently slower than others in reading data.

Which of the following could be contributing factors to this performance discrepancy? Select all that apply.

Network congestion between the slower nodes and the storage system.

Uneven data distribution across the storage nodes.

Different CPU architectures on the nodes.

Insufficient RAM on the slower nodes for caching data.

Variations in the speed of the local temporary storage (e.g., /tmp) used for intermediate files.

55. Consider a scenario where you’re using GPUDirect Storage to enable direct memory access between GPUs and NVMe drives. You observe that while GPUDirect Storage is enabled, you’re not seeing the expected performance gains.

What are potential reasons and configurations you should check to ensure optimal GPUDirect Storage performance? Select all that apply.

Verify that the NVMe drives are properly configured in a RAID 0 configuration.

Ensure that the NVMe drives are connected to the system via PCle Gen4 or Gen5.

Confirm that the CUDA driver version is compatible with GPIJDirect Storage.

Check if the file system supports direct I/O (e.g., using ‘directio’ mount option).

Disable CPU-side caching to force all I/O operations to go directly to the GPU memory.

56. You’re working with a large dataset of microscopy images stored as individual TIFF files. The images are accessed randomly during a training job. The current storage solution is a single HDD. You’re tasked with improving data loading performance.

Which of the following storage optimizations would provide the GREATEST performance improvement in this specific scenario?

Implementing data deduplication on the storage volume.

Migrating the data to a large, sequential HD

Replacing the HDD with a RAID 5 array of HDDs.

Replacing the HDD with a single NVMe SS

Compressing the TIFF files using a lossless compression algorithm.

57. A large A1 model is training using a dataset stored on a network-attached storage (NAS) device. The data transfer speeds are significantly lower than expected. After initial troubleshooting, you discover that the MTU (Maximum Transmission Unit) size on the network interfaces of the training server and the NAS device are mismatched. The server is configured with an MTIJ of 1500, while the NAS device is configured with an MTU of 9000 (Jumbo Frames).

What is the MOST likely consequence of this MTU mismatch, and what action should you take?

Data packets will be fragmented, leading to increased overhead and reduced performance. Configure both the server and the NAS device to use the same MTU size (either 1500 or 9000).

The connection between the server and the NAS device will be unreliable, resulting in data corruption. Increase the MTU size on both devices to the maximum supported value.

The server will be unable to communicate with the NAS device. Reduce the MTU size on the server to match the MTU size of the NAS device.

The data transfer will be limited to the lowest common MTU size, but there will be no significant performance impact. No action is required.

Data packets will be retransmitted, increasing the latency but still getting the full throughput. Configure the server to use Path MTU Discovery (PMTUD).

58. You are troubleshooting slow I/O performance in a deep learning training environment utilizing BeeGFS parallel file system. You suspect the metadata operations are bottlenecking the training process.

How can you optimize metadata handling in BeeGFS to potentially improve performance?

Increase the number of storage targets (OSTs) to distribute the data across more devices.

Implement data striping across multiple OSTs.

Increase the number of metadata servers (MDSs) and distribute the metadata load across them.

Enable client-side caching of metadata on the training nodes.

Configure BeeGFS to use a different network protocol with lower overhead.

59. You are designing a storage solution for a new AI inference cluster that requires extremely low latency for model serving.

Which storage technology and configuration would be MOST suitable to meet this stringent latency requirement?

A distributed file system deployed on spinning HDDs with a large read-ahead cache.

NVMe-oF (NVMe over Fabrics) using RDMA over Converged Ethernet (RoCE) connected to a cluster of NVMe drives.

A software-defined storage (SDS) solution running on commodity hardware with SATA SSDs.

Amazon S3 object storage accessed over a high-bandwidth internet connection.

A traditional Fibre Channel SAN with a dedicated storage array.

60. You have a large dataset stored on a BeeGFS file system. The training job is single node and uses data augmentation to generate more data on the fly. The data augmentation process is CPU-bound, but you notice that the GPU is underutilized due to the training data not being fed to the GPU fast enough.

How can you reduce the load on the CPU and improve the overall training throughput?

Move the training data to a local NVMe drive on the training node.

Increase the number of BeeGFS metadata servers (MDSs) to improve metadata performance.

Implement asynchronous 1/0 in the data loading pipeline using a library like NVIDIA DALI to offload data processing tasks from the CPU to the GP

Decrease the batch size of the training job to reduce the amount of data being processed at each iteration.

Enable data compression on the BeeGFS file system to reduce the amount of data being transferred over the network.

Page 7 of 10

61. You are configuring an InfiniBand subnet with multiple switches. You need to ensure that traffic between two specific nodes always takes the shortest path, bypassing a potentially congested link.

Which of the following approaches is MOST effective for achieving this using InfiniBand’s routing capabilities?

Rely solely on the Subnet Manager’s (SM) default path computation algorithm (e.g., Min Hop) without any modifications.

Use static routing by manually configuring forwarding tables on each switch along the desired path. This involves specifying DLID-to-Port mappings.

Implement Quality of Service (QOS) to prioritize the traffic between the two nodes, hoping that this will influence the path selection.

Utilize the ibroute command or similar tool to inject a static route between the nodes, forcing traffic to follow a specific path identified by LID and port number.

Decrease the MTIJ size on the potential congested link.

62. You are using NVIDIA Spectrum-X switches in your A1 infrastructure. You observe high latency between two GPU servers during a large distributed training job. After analyzing the switch telemetry, you suspect a suboptimal routing path is contributing to the problem.

Which of the following methods offers the MOST granular control for influencing traffic flow within the Spectrum-X fabric to mitigate this?

Adjust the Equal-Cost Multi-Path (ECMP) hashing algorithm globally on all switches.

Configure QOS (Quality of Service) policies to prioritize traffic from the high-latency GPU servers.

Implement Adaptive Routing (AR) or Dynamic Load Balancing (DLB) features available on the Spectrum-X switches to dynamically adjust paths based on network conditions.

Manually configure static routes on the Spectrum-X switches to force traffic between the GPU servers along a specific path.

Disable IPv6 to simplify routing decisions.

63. In an InfiniBand fabric, what is the primary role of the Subnet Manager (SM) with respect to routing?

To forward packets based on destination IP addresses, similar to a traditional IP router.

To discover the network topology, calculate routing paths, and program the forwarding tables (LID tables) in the switches.

To monitor the network for congestion and dynamically adjust packet priorities using Quality of Service (QOS) mechanisms.

To provide a command-line interface for users to manually configure routing tables on each InfiniBand switch.

To act as a firewall, blocking unauthorized traffic based on pre-defined rules.

64. Consider the following ‘ibroute’ command used on an InfiniBand host: ‘ibroute add dest Oxla dev ib0’.

What is the MOST likely purpose of this command?

To add a default route for all traffic destined outside the InfiniBand subnet.

To create a static route for traffic destined to LID Ox1a, using the InfiniBand interface ib0.

To configure the MTU size on the ib0 interface to Ox1a bytes.

To disable routing on the ib0 interface.

To configure a static route for traffic destined to IP address Ox1a, using the InfiniBand interface ib0.

65. You are troubleshooting a network performance issue in your NVIDIA Spectrum-X based A1 cluster. You suspect that the Equal-Cost Multi-Path (ECMP) hashing algorithm is not distributing traffic evenly across available paths, leading to congestion on some links.

Which of the following methods would be MOST effective for verifying and addressing this issue?

Use ‘ping’ or ‘traceroute’ to analyze the paths taken by packets between the affected nodes. If they always take the same path, ECMP is likely not working correctly.

Use switch telemetry tools (e.g., NVIDIA What’s Up Gold, Mellanox NEO, or similar) to monitor link utilization across all available paths between the nodes. Look for significant imbalances in traffic volume.

Restart the switches to force the ECMP hashing algorithm to recalculate paths.

Disable ECMP entirely and rely solely on static routing.

Reduce the TCP window size.

66. An InfiniBand fabric is experiencing intermittent packet loss between two high-performance compute nodes. You suspect a faulty cable or connector.

Besides physically inspecting the cables, what software-based tools or techniques can you employ to diagnose potential link errors contributing to this packet loss?

Use ‘ibdiagnet’ to perform a comprehensive fabric analysis, including link integrity checks and error detection.

Monitor the port counters on the InfiniBand switches connected to the compute nodes. Look for excessive CRC errors, symbol errors, or other link-related error counts.

Run ‘ipeff or ‘ibperf between the two compute nodes and analyze the reported packet loss rate. Correlate this with the error counters on the switches.

All of the above

Disable port mirroring.

67. You are managing an A1 infrastructure based on NVIDIA Spectrum-X switches. A new application requires strict Quality of Service (QOS) guarantees for its traffic. Specifically, you need to ensure that this application’s traffic receives preferential treatment and minimal latency.

What combination of Spectrum-X features and configurations would be MOST effective in achieving this?

Configure DiffServ Code Point (DSCP) marking on the application’s traffic, map these DSCP values to specific traffic classes within the Spectrum-X switch, and configure Weighted Fair Queueing (WFQ) or Strict Priority Queueing on the egress ports.

Increase the MTIJ size on all interfaces to reduce packet fragmentation and overall latency.

Disable Adaptive Routing (AR) to ensure that traffic always takes the shortest path.

Use VLAN tagging to isolate the application’s traffic into a separate virtual network.

Enable broadcast storm protection.

68. Which of the following statements are true regarding the use of Congestion Management (CM) and Congestion Avoidance (CA) techniques within an InfiniBand fabric using NVIDIA technology? (Select TWO)

CM/CA mechanisms are primarily implemented at the IP layer and are independent of the InfiniBand transport layer.

CM aims to reduce the severity of congestion once it has already occurred, while CA aims to prevent congestion from happening in the first place.

InfiniBand’s Explicit Congestion Notification (ECN) is a CA mechanism that allows switches to signal congestion to endpoints before packet loss occurs.

CM/CA are not relevant in InfiniBand fabrics because InfiniBand’s lossless nature guarantees that no packets will ever be dropped due to congestion.

CM can include techniques like rate limiting to throttle traffic flows when congestion is detected.

69. In a large-scale InfiniBand fabric, you need to implement a mechanism to prioritize traffic for a specific application that requires low latency and high bandwidth. You want to leverage Quality of Service (QOS) to achieve this.

Which of the following steps are essential to properly configure QOS in this scenario? (Select THREE)

Configure VLAN tagging on the application’s traffic to isolate it from other traffic.

Map the application’s traffic to a specific traffic class with appropriate priority settings within the InfiniBand switches.

Configure Weighted Fair Queueing (WFQ) or Strict Priority Queueing on the egress ports of the InfiniBand switches to prioritize the application’s traffic class.

Disable Adaptive Routing (AR) to ensure that the application’s traffic always takes the shortest path.

Mark the application’s traffic with appropriate DiffServ Code Point (DSCP) values.

70. You’re designing a new InfiniBand network for a distributed deep learning workload. The workload consists of a mix of large-message all- to-all communication and small-message parameter synchronization.

Considering the different traffic patterns, what routing strategy would MOST effectively minimize latency and maximize bandwidth utilization across the fabric?

Rely solely on the default Subnet Manager (SM) with a Min Hop path selection algorithm.

Implement a static routing scheme with manually configured forwarding tables on each switch.

Utilize a combination of Adaptive Routing (AR) to handle dynamic traffic patterns and Quality of Service (QOS) to prioritize small-message parameter synchronization.

Implement a purely deterministic routing scheme, disabling all adaptive routing features.

Disable multicast.

Page 8 of 10

71. You are experiencing link flapping (frequent up/down transitions) on several InfiniBand links in your AI infrastructure. This is causing intermittent connectivity issues and performance degradation.

What are the MOST likely causes of this issue, and what steps should you take to troubleshoot and resolve it? (Select TWO)

Incorrect MTU (Maximum Transmission Unit) configuration on the affected interfaces.

Faulty or damaged cables, connectors, or transceivers.

Software bugs in the operating system or InfiniBand drivers.

Mismatched link speeds or duplex settings between connected devices.

Excessive broadcast traffic causing congestion.

72. Which of the following are key benefits of using NVIDIA Spectrum-X switches in an A1 infrastructure compared to traditional Ethernet switches? (Select THREE)

Lower cost per port.

Support for RoCE (RDMA over Converged Ethernet) and InfiniBand protocols, enabling high-bandwidth, low-latency communication.

Advanced telemetry and monitoring capabilities for network performance optimization.

Hardware-based acceleration for collective communication operations used in distributed A1 training.

Native support for IPv6.

73. Which of the following are key benefits of using NVIDIA NVLink� Switch in a multi-GPU server setup for AI and deep learning workloads?

Increased GPU-to-GPIJ communication bandwidth.

Reduced latency in inter-GPU data transfers.

Simplified GPU resource management.

Support for larger GPU memory pools than a single server can physically accommodate.

Enhanced security features compared to PCle based interconnections.

74. You are tasked with configuring an NVIDIA NVLink� Switch system. After physically connecting the GPUs and the switch, what is the typical first step in the software configuration process?

Installing the latest NVIDIA drivers on all connected GPUs.

Configuring the system BIOS to enable NVLink support.

Updating the firmware of the NVLink Switch.

Installing the NVLink Switch management software.

Running a memory bandwidth test between all connected GPUs.

75. Which command-line tool is typically used to monitor the status and performance of an NVIDIA NVLink� Switch?

nvidia-smi

nvswitch-cli

ibstat

rocminfo

Ispci

76. During NVLink Switch configuration, you encounter issues where certain GPUs are not being recognized by the system.

Which of the following troubleshooting steps are most likely to resolve this problem?

Verify that all NVLink cables are securely connected and properly seated.

Check the system BIOS settings to ensure that NVLink is enabled and configured correctly.

Ensure that the NVLink Switch firmware is compatible with the installed GPUs.

Reinstall the operating system.

Check the Power supply for enough capacity and stability.

77. When setting up a multi-server, multi-GPU environment using NVLink switches, what is the primary consideration when planning the network topology for optimal performance?

Minimizing the number of hops between GPUs that need to communicate frequently.

Maximizing the distance between servers to improve cooling.

Using a star topology for simplified management.

Ensuring all servers are on the same subnet for ease of configuration.

Placing servers near the network’s edge to reduce latency.

78. What is the role of GPUDirect RDMA in an NVLink Switch-based system, and how does it improve performance?

It allows GPUs to directly access each other’s memory without involving the CPIJ, reducing latency and CPU overhead.

It provides a mechanism for GPUs to offload compute-intensive tasks to the CPU, improving overall system throughput.

It enables direct communication between GPUs and storage devices, bypassing the network interface.

It facilitates the virtualization of GPUs, allowing multiple virtual machines to share a single physical GPI

It encrypts data transmitted between GPUs, enhancing security.

79. Consider a scenario where you are using NCCL (NVIDIA Collective Communications Library) for multi-GPU training across multiple servers connected via NVLink switches.

Which NCCL environment variable would you use to specify the network interface to be used for communication?

NCCL PORT

NCCL SOCKET IFNAME

NCCL NET INTERFACE

NCCL 1B HCA

NCCL COMM ID

80. You observe high latency and low bandwidth between two GPUs connected via an NVLink switch. You suspect a problem with the NVLink link itself.

Which of the following methods would be the most effective in diagnosing the physical NVLink link health?

Using ‘iperf3’ to measure network throughput between the servers.

Running a CUDA-aware memory bandwidth test specifically designed for NVLink.

Examining system logs for NVLink-related error messages.

Using ‘ping’ to check network connectivity between the servers.

Physically inspecting the NVLink cables for damage.

Page 9 of 10

81. In a distributed training environment with NVLink switches, you need to optimize the data transfer between GPUs on different servers.

Which strategy is most likely to minimize the impact of inter-server latency on the overall training time?

Increasing the batch size to amortize the cost of data transfers.

Using asynchronous data transfers with overlapping computation.

Compressing the data before transferring it over the network.

Using a centralized parameter server architecture.

Switching to a synchronous SGD (Stochastic Gradient Descent) algorithm.

82. Given the following ‘nvswitch-cli’ output, what does the ‘Link Speed’ indicate, and what potential bottleneck might a low ‘Link Speed’ suggest?

It indicates the effective bandwidth of the NVLink connection; a low value suggests a potential cable issue or misconfiguration.

It indicates the clock speed of the GPU memory; a low value suggests a memory bottleneck.

It indicates the PCle generation supported by the GPIJ; a low value suggests an outdated GP

It indicates the NVLink protocol version; a low value suggests firmware incompatibility.

It indicates the power consumption of the NVLink switch; a high value suggests overheating issues.

83. You are implementing a distributed deep learning training setup using multiple servers connected via NVLink switches. You want to ensure optimal utilization of the NVLink interconnect.

Which of the following strategies would be MOST effective in achieving this goal?

Configure NCCL to use GPUDirect RDMA for inter-GPU communication across servers.

Use a standard TCP/IP socket connection for inter-GPU communication across servers.

Implement a data compression algorithm that can be processed by the CPU before sending data over NVLink.

Disable peer-to-peer GPU memory access within each server to avoid contention.

Increase the batch size to reduce the frequency of inter-GPU communication.

84. You are deploying a new NVLink Switch based cluster. The GPUs are installed in different servers, but need to be configured to utilize

NVLink interconnect.

Which of the following should be performed during the installation phase to confirm correct configuration?

Run NCCL tests to verify the GPU-to-GPU bandwidth and latency between servers.

Verify that GPUDirect RDMA is enabled and functioning correctly.

Check that the ‘nvidia-sm’ command shows the correct NVLink topology.

Run standard TCP/IP network bandwidth tests to check inter-server communication.

All the GPU’s are in the same IP subnet

85. You are configuring a Mellanox InfiniBand network for a DGXAIOO cluster.

What is the RECOMMENDED subnet manager for a large, high-performance A1 training environment, and why?

OpenSM, because it’s the default and easiest to configure.

UFM (Unified Fabric Manager), because it provides advanced management, monitoring, and optimization capabilities.

IBA management tools that ship with the OS (e.g., ‘ibnetdiscover’).

Any subnet manager; the performance difference is negligible.

A custom-built subnet manager using the InfiniBand verbs AP

86. You are setting up network fabric ports for hosts in an NVIDIA-Certified Professional A1 Infrastructure (NCP-AII) environment. You need to configure Jumbo Frames to improve network throughput.

What is the typical MTU (Maximum Transmission Unit) size you would set on the network interfaces and switches, and why?

1500 bytes, as it’s the default and compatible with most networks.

9000 bytes, also known as Jumbo Frames, reduces overhead and improves throughput for large data transfers.

65535 bytes, the theoretical maximum MTU size, for maximum performance.

576 bytes, the minimum MTU size required by IPv4.

Any MTU size between 1500 and 9000 bytes; the specific value doesn’t matter.

87. Which of the following are valid methods for verifying the health and connectivity of InfiniBand links in an NCP-AII environment? (Select TWO)

Using ‘ping’ to test basic IP connectivity over the InfiniBand interface.

Using ‘ibstat’ to check the link state, physical state, and other relevant parameters of InfiniBand ports.

Using ‘netstat’ to check TCP connections.

Using ‘sminfo’ to query the Subnet Manager for network topology and status information.

Checking the system logs ( ‘ /var/log/messages’ or equivalent) for any InfiniBand-related error messages.

88. You’re configuring a RoCEv2 network for your AI infrastructure.

Which UDP port number range is commonly used for RoCEv2 traffic, and why is it important to be aware of this?

0-1023, because these are well-known ports.

4791, which is reserved for VXLA

49152-65535, the dynamic/private port range, to avoid conflicts with other services.

1024-49151, the registered port range, for general application use.

Any UDP port number can be used without issue.

89. You are configuring network fabric ports for NVIDIA GPUs in a server. The GPUs are connected to the network via PCIe.

What is the primary factor that determines the maximum achievable bandwidth between the GPUs and the network?

The clock speed of the CP

The amount of system RA

The PCIe generation and number of lanes connecting the GPUs to the network adapter (e.g., PCIe 4.0 x16).

The speed of the system’s hard drives or SSDs.

The color of the Ethernet cables.

90. You are troubleshooting a network performance issue in your NCP-AII environment.

After running ‘ibstat’ on a host, you see the following output for one of the InfiniBand ports:

What does the ‘LMC: 0’ indicate, and what are the implications for network performance?

LMC: 0 indicates that the link is down and not functioning correctly.

LMC: 0 indicates that Link Aggregation (LAG) is not enabled on this port, meaning only a single link is being used for communication.

LMC: 0 indicates the port is operating at the lowest possible speed.

LMC: 0 indicates that the Subnet Manager is not running correctly.

LMC: 0 is the default and expected value; it has no impact on performance.

Page 10 of 10

91. You are deploying a multi-tenant AI infrastructure where different users or groups have isolated network environments using VXLAN.

Which of the following is the MOST important consideration when configuring the VTEPs (VXLAN Tunnel Endpoints) on the hosts to ensure proper network isolation and performance?

Using the default MTU size of 1500 bytes for VXLAN traffic.

Ensuring that each tenant has a unique VXLAN Network Identifier (VNI) to isolate their traffic.

Using the same IP address for all VTEPs to simplify routing.

Disabling multicast routing to prevent broadcast traffic.

Using the same VNI for all tenants to maximize network utilization.

92. You are configuring a network bridge on a Linux host that will connect multiple physical network interfaces to a virtual machine. You need to ensure that the virtual machine receives an IP address via DHCP.

Which of the following is the correct command sequence to create the bridge interface ‘br0’, add physical interfaces ‘eth0’ and ‘eth1’ to it, and bring up the bridge interface? Assume the required packages are installed. Consider using ‘ip’ command.

A )

B )

C )

D )

E )

Option A

Option B

Option C

Option D

Option E

93. You are planning the network infrastructure for a DGX SuperPOD. You need to ensure that the network fabric can handle the high bandwidth and low latency requirements of A1 training workloads.

Which network technology is the RECOMMENDED choice for interconnecting the DGX nodes within the SuperPOD, and why?

Gigabit Ethernet, because it’s widely available and inexpensive.

10 Gigabit Ethernet, for a balance between cost and performance.

InfiniBand, due to its high bandwidth, low latency, and RDMA support.

Wi-Fi 6, for wireless connectivity and flexibility.

Token Ring, because it’s a reliable and deterministic networking protocol.

94. You are configuring a switch port connected to a host in an NCP-AII environment. The host is running RoCEv2.

To optimize performance and prevent packet loss, which flow control mechanism should you enable on the switch port?

None; flow control is not needed with RoCEv2.

TCP flow control.

Priority Flow Control (PFC) or 802.1 Qbb, specifically for the traffic class associated with RoCEv2.

Simple Network Management Protocol (SNMP).

Spanning Tree Protocol (STP).

95. You are tasked with setting up network fabric ports to connect several servers, each with multiple NVIDIA GPUs, to an InfiniBand switch. Each server has two ConnectX-6 adapters.

What is the best strategy to maximize bandwidth and redundancy between the servers and the InfiniBand fabric?

Connect only one adapter from each server to the switch to minimize cable clutter.

Connect both adapters from each server to the same switch, but do not configure link aggregation.

Connect both adapters from each server to the same switch and configure link aggregation (LACP or static LAG) on both the server and the switch.

Connect one adapter from each server to one switch, and the second adapter to a different switch, without link aggregation.

Connect one adapter from each server to one switch, and the second adapter to a different switch, and configure multi-pathing on the servers.

NCP-AII NVIDIA Certified Professional AI Infrastructure Exam Free Dumps

NCP-AII NVIDIA Certified Professional AI Infrastructure Exam Free Dumps

Share this post

Leave a Reply Cancel reply

Related Posts

NCA-AIIO AI Infrastructure and Operations Exam Free Dumps

NCA-GENL NVIDIA Generative AI LLMs Exam Free Dumps

NCP-AIN NVIDIA-Certified Professional AI Networking Exam Free Dumps

NCP-AIO NVIDIA Certified Professional AI Operations Exam Free Dumps