NVIDIA NCP-AII 試験概要:

NVIDIA NCP-AII 認定試験の出題範囲：

参照：https://www.nvidia.com/en-us/learn/certification/ai-infrastructure-professional/

NVIDIA AI Infrastructure 認定 NCP-AII 試験問題：

問題 #1
After configuring HA, the administrator runs cmsh status and notices the secondary head node reports mysql [FAIL]. What is the most likely cause?

A. The BCM license expired after HA configuration.
B. Network connectivity issues between the primary and secondary head nodes.
C. The cluster nodes are powered on during the HA configuration.
D. The secondary head node lacks NVIDIA GPU drivers.

問題 #2
An infrastructure engineer runs a NCCL burn-in on an eight-node GPU cluster. Over a 12-hour period, all GPUs are tested with repeated all-reduce collectives. Monitoring tools show the following observations:
- Aggregate bandwidth remains within 5% of documented reference for the hardware on every run.
- No errors or timeouts are reported in NCCL logs.
- On three occasions, one GPU logged single-run bandwidth dips of 15-
20% compared to its norm, but performance recovered on the next run and stayed stable afterward. System logs show no hardware or drive errors.
- Two minor NCCL WARN-level messages about "unexpected latency spike"
appear in system logs for separate nodes, but could not be reproduced.
Which conclusion is the best strategy before releasing the cluster to production?

A. Recommend proactive maintenance, because any bandwidth drop-even if transient and unreproducible - shows the burn-in failed; clusters must not show performance variance above
10% for any GPU even once.
B. Approve for AI workload use, but flag affected nodes for manual exclusion from distributed training jobs, as nodes showing any anomaly should be isolated whenever possible.
C. Proceed, since all bandwidth targets are met, issues were transient and self-resolved, and there are no persistent errors or timeouts across repeated burn-ins.

問題 #3
You are training a deep neural network using NCCL to coordinate communication across four GPUs in a single node. During early performance testing, you notice inconsistent scaling and longer-than-expected training times, even though all GPUs are being used. Which strategy would most effectively improve NCCL efficiency and collective operation performance in this setting?

A. Adjust the batch size so that each GPU receives an equalized portion of the batch, ensuring all GPUs process similar workloads end communication is evenly distributed.
B. Disable automatic load balancing so that the deep learning framework can dynamically assign samples to any GPU available during each iteration.
C. Assign the largest possible workload to the first GPU to maximize its utilization, and allow the remaining GPUs to process smaller or variable batch sizes as needed.
D. Increase the communication frequency between GPUS allowing workloads to be unevenly split, so that synchronization is more frequent and model updates happen faster.

問題 #4
A financial services firm is deploying an AI model for fraud detection that requires rapid inference and data retrieval across multiple sites. Which feature should their storage system prioritize?

A. Tape backup systems.
B. Multi-protocol data access with low latency.
C. High capacity with moderate speed.
D. Low-cost HDD solutions.

問題 #5
Which of the following steps are essential components of a recommended DGX cluster installation procedure? (Choose two.)

A. Configure networking by validating all interfaces on each node - ensuring proper InfiniBand/Ethernet connectivity - prior to installing cluster software.
B. Complete application containerization, run distributed jobs, validate node health or storage availability.
C. Install Slurm on the head node and then configure the compute nodes' default OS images.
D. Group nodes by function during initial setup and assign them to relevant categories in the cluster management tool.

解説：