Describe Solutions to Maximize Performance in AI Workloads CERTIFIED-AI-EXPERT Practice Questions (Page 5)

Questions 21–25

1. 21application · medium

   A computer vision team trains a model on a large image dataset. Each epoch takes 6 hours, and GPU utilization is only 50%. Profiling shows the GPUs are idle about 40% of the time waiting for the next batch. The preprocessing pipeline (resize, crop, normalize) runs on CPU workers and takes 300 ms per image. The team has 8 GPUs and 64 CPU cores. Which change would most effectively close the GPU idle gap?

   • APrefetch the next batch to GPU memory while the current batch is being processed.
   • BReduce the number of CPU workers to free up cores for the GPU.
   • CSwitch to a smaller batch size to process images faster.
   • DIncrease the image resolution to make the model more accurate.
   Reveal Solution Discussion Select an answer first
2. 22foundation · medium

   Which scenario is best suited for vertical scaling of compute resources?

   • AA workload that is already running on a cluster of many small nodes.
   • BA workload that requires more memory or a faster CPU on a single node.
   • CA workload that can be parallelized across many nodes.
   • DA workload that needs to handle variable traffic patterns.
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   A research team is training a large language model on a 100 Gbps cluster. GPU utilization is consistently below 40%, and the training loop shows long gaps between data batches. Profiling reveals that the data loader spends most of its time waiting on file reads from a shared NFS volume, while CPU utilization on the worker nodes is low. The team has already increased the number of data-loading worker processes. Which action would most directly address the identified bottleneck?

   • AMove the training dataset to a high-throughput parallel file system or NVMe-backed local NVMe drives on each worker node.
   • BEnable gradient compression to reduce communication overhead between GPUs.
   • CSwitch from data parallelism to model parallelism to reduce the amount of data each GPU needs.
   • DIncrease the batch size to reduce the number of data-loading iterations per epoch.
   Reveal Solution Discussion Select an answer first
5. 24foundation · medium

   Which prefetching strategy is most effective for hiding I/O latency in a sequential data pipeline?

   • APrefetching only after the current batch is fully consumed.
   • BPrefetching all data into memory at the start of training.
   • CPrefetching random data from the entire dataset.
   • DPrefetching the next batch while the current batch is being processed.
   Reveal Solution Discussion Select an answer first
6. 25application · medium

   A team is running data-parallel training of a large model on 16 GPUs connected via 10 Gbps Ethernet. Training is slow, and profiling shows that GPUs spend 50% of the time waiting for gradient synchronization. The model has 1 billion parameters. Which change would most directly reduce the communication bottleneck?

   • ASwitch to model parallelism to split the model across GPUs.
   • BUpgrade the interconnect to InfiniBand or 100 Gbps RoCE.
   • CUse a single GPU to avoid communication entirely.
   • DReduce the batch size to shrink gradient sizes.
   Reveal Solution Discussion Select an answer first