NCP-AII Practice Questions

A.  

Run a deep learning workload to stress test the GPUs and check whether the issue persists.

B.  

Check the NVIDIA System Management Interface (nvidia-smi) for GPU status and temperatures.

C.  

Power drain then restart the DGX and check if the performance degradation resolves.

D.  

Increase the fan speed to maximum and check whether the performance improves.

A.  

Local SSD cache allows users to increase the number of NFS threads on the server without impacting storage reliability.

B.  

Using local SSD cache in RAID-0 enables direct GPU access to files without host CPU involvement, further boosting performance.

C.  

Local SSD cache in RAID-0 is necessary to provide redundancy in case one of the drives fails during long training runs.

D.  

A local SSD cache in RAID-0 ensures that most training data is read only once from the network, significantly reducing NFS traffic.

A.  

tcpdump

B.  

ib_write_lat

C.  

ibdiagnet

D.  

perfmon

A.  

Update the GPU driver leaving the DOCA and OFED drivers unchanged as long as they are detecting the hardware properly.

B.  

Validate the driver version post-install since the fresh install will overwrite the legacy drivers.

C.  

Keep the older driver running alongside the new version in case you need to roll back the upgrade.

D.  

Uninstall all existing GPU and DOCA-related drivers and associated kernel modules before the new install.

A.  

Use Power Usage Effectiveness as the primary metric while supplementing it with additional measures of useful work done per unit of energy.

B.  

Use watts used as the primary measure of efficiency, as it accurately reflects the power input at any given time.

C.  

Develop benchmarks tailored to specific workloads, such as MLPerf for AI applications, to better understand energy use in real-world scenarios.

D.  

Focus on integrating kilowatt-hours into existing metrics to better reflect the actual energy used for productive work.

A.  

mlxconfig -d q

B.  

ethtool -S

C.  

mlxlink -d -c -e

D.  

mstflint -d q full

A.  

lspci | grep NVIDIA

B.  

nvidia-smi

C.  

nvidia-gpu-status

D.  

iblinkinfo

A.  

Use NCCL_TESTS_SPLIT="DIV 8" with point-to-point tests

B.  

Use all_reduce_perf -b 8 -e 16G -f 2 -g 8 with NCCL_TESTS_SPLIT="AND 0x1"

C.  

Use reduce_scatter_perf -b 8 -e 16G -f 2 -g 4

D.  

Use all_reduce_perf -b 8 -e 16G -f 2 -g 8 without splits

A.  

Lane power variance < 3dB across all transceivers.

B.  

Transceiver model matching QSFP-DD specifications.

C.  

Temperature fluctuations > 5°C during validation.

D.  

Effective BER > 1.5E-254 during a <6-hour monitoring window.

A.  

Anti-ESD strap

B.  

Gloves

C.  

Protective film

D.  

Electric screwdriver