Describe Solutions to Maximize Performance in AI Workloads CERTIFIED-AI-EXPERT Practice Questions (Page 2)

Questions 6–10

1. 6application · medium

   A team is training a transformer model on 4 NVIDIA A100 GPUs. GPU utilization is only 30%, and the model fits in memory. Profiling shows the forward and backward passes are fast, but the GPUs spend most of the time idle between steps. The team suspects the small batch size is the cause. Which change would most directly improve GPU utilization?

   • AEnable gradient accumulation with a smaller effective batch.
   • BReduce the batch size to speed up each iteration.
   • CSwitch to a smaller model to reduce memory usage.
   • DIncrease the batch size until GPU memory is nearly full.
   Reveal Solution Discussion Select an answer first
2. 7application · medium

   A team trains a model on NVIDIA GPUs that support mixed precision (FP16). The current training uses FP32 and takes 8 hours. The team wants to reduce training time without changing the model architecture. Which change would most directly reduce training time?

   • AIncrease the batch size to use more memory.
   • BReduce the number of GPUs to simplify synchronization.
   • CSwitch to a larger model to use the GPU more efficiently.
   • DEnable mixed precision training with FP16 and FP32 master weights.
   Reveal Solution Discussion Select an answer first
4. 8foundation · medium

   Which file system feature is particularly beneficial for AI workloads that read many small files?

   • ALarge block size
   • BCompression enabled by default
   • CHigh metadata throughput
   • DSnapshots for data protection
   Reveal Solution Discussion Select an answer first
5. 9foundation · medium

   Which tool is commonly used to profile GPU utilization and identify bottlenecks in deep learning training?

   • ANVIDIA Nsight Systems
   • Bgprof
   • Chtop
   • DWireshark
   Reveal Solution Discussion Select an answer first
6. 10foundation · medium

   What is the primary advantage of data parallelism over model parallelism?

   • AIt is easier to implement and scales well with many GPUs.
   • BIt eliminates the need for a high-speed interconnect.
   • CIt allows training models that are larger than any single GPU's memory.
   • DIt reduces the need for gradient synchronization.
   Reveal Solution Discussion Select an answer first