Logo ExamsQA
Exit
Official Bank 0/70

NCP-AII AI Infrastructure professional (NCP-AII) - Nvidia Exam Questions

Last updated on July 22, 2026

97% Exam Compliance
70 Total Questions
1
Question
For an NVIDIA Enterprise AI Factory with 256 GPUs, which storage solution characteristic is most
critical to validate during scaling tests?
Options
A Single-node write performance during idle clusters.
B RAID rebuild times under disk failure.
C Maximum 4K random read IOPS exceeding 1 million.
D Consistent per-node throughput >8 GiB/s.
Discussion (0 comments)

to join the discussion

Community Discussion

No discussions yet. Be the first to ask!

2
Question
An engineer needs to validate NVLink Switch functionality on a DGX H100 system with 8 GPUs.

Which NCCL command verifies intra-node NVLink bandwidth?
Options
A all_reduce_perf -b 8 -e 16G -f 2 -g 4 with NCCL_TESTS_SPLIT="MOD 2"
B broadcast_perf -b 8 -e 16G -f 2 -g 8 without split configuration
C all_reduce_perf -b 8 -e 16G -f 2 -g 8 with NCCL_TESTS_SPLIT="OR 0x7"
D all_reduce_perf -b 8 -e 16G -f 2 -g 1 repeated 8 times
Discussion (0 comments)

to join the discussion

Community Discussion

No discussions yet. Be the first to ask!

3
Question
When updating the firmware on an NVLink switch transceiver, how can an engineer apply new firmware without interrupting the network?
Options
A Physically disconnect and reconnect the transceiver.
B flint -d -lid 27 --linkx --linkx_auto_update --activate
C nv action reboot system to force immediate activation.
D mlxfwreset -d -lid 27 reset --yes to reset the transceiver
Discussion (0 comments)

to join the discussion

Community Discussion

No discussions yet. Be the first to ask!

4
Question
An InfiniBand administrator needs to run performance benchmarks on new devices added to the fabric. What tool should be used to check the latency?
Options
A ib_write_lat
B ibdiagnet
C perfmon
D tcpdump
Discussion (0 comments)

to join the discussion

Community Discussion

No discussions yet. Be the first to ask!

5
Question
During a 48-hour NeMo question-answering model burn-in test, GPU memory errors occur when processing large datasets. Which configuration strategy prevents Out-of-Memory (OOM) errors while maintaining processing efficiency?
Options
A Switch from FP16 to FP32 precision for numerical stability.
B Disable add_filename for Parquet files to reduce metadata.
C Increase files_per_partition to 1000 for larger batch processing.
D Set blocksize="1GB" for data loading and enable RMM asynchronous allocation.
Discussion (0 comments)

to join the discussion

Community Discussion

No discussions yet. Be the first to ask!

Finish Practice?

Are you sure you want to finish? This will end your practice session.