Describe Solutions to Maximize Performance in AI Workloads CERTIFIED-AI-EXPERT Practice Questions (Page 4)

Questions 16–20

1. 16foundation · medium

   Which scaling strategy involves adding more GPUs to a single training job?

   • ADiagonal scaling
   • BElastic scaling
   • CVertical scaling
   • DHorizontal scaling
   Reveal Solution Discussion Select an answer first
2. 17foundation · medium

   What is the primary benefit of performing data augmentation on the fly (during training) rather than pre-generating an augmented dataset?

   • AIt reduces the number of training steps required.
   • BIt guarantees the model will achieve higher accuracy.
   • CIt reduces the amount of storage required for the training dataset.
   • DIt eliminates the need for a validation set.
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A team runs a training job that is memory-bound: the model and activations use 80% of the 32 GB GPU memory. The team wants to increase the batch size to improve GPU utilization, but the batch does not fit in memory. Which approach would allow a larger effective batch size without changing the model?

   • AReduce the batch size to fit in memory.
   • BIncrease the learning rate to compensate for the small batch.
   • CSwitch to a smaller model to free memory.
   • DUse gradient accumulation to simulate a larger batch size.
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   A team trains a model on a dataset stored in a remote object store. Each epoch, the data loader fetches images over the network, and GPU utilization is low because of network latency. The team has enough local NVMe storage to hold the entire dataset. Which strategy would most effectively reduce the data-access latency?

   • AIncrease the number of data-loader workers to fetch more images in parallel.
   • BUse a slower network to reduce congestion.
   • CReduce the image resolution to shrink the download size.
   • DDownload the dataset to local NVMe storage once and read from local disk during training.
   Reveal Solution Discussion Select an answer first
6. 20expert · hard

   A team trains a model on a dataset of high-resolution medical images. The preprocessing pipeline (resize, normalize, augment) is CPU-intensive and takes 500 ms per image. The team has 4 GPUs and 32 CPU cores. GPU utilization is only 40%, and the GPUs are idle waiting for preprocessed batches. The team has already tried increasing the number of data-loader workers, but CPU usage is already at 100%. Which change would most effectively improve GPU utilization?

   • AAdd more CPU cores to the worker nodes.
   • BIncrease the batch size to reduce the number of preprocessing calls.
   • CMove the preprocessing to the GPU using DALI or GPU-accelerated augmentation.
   • DReduce the image resolution to speed up preprocessing.
   Reveal Solution Discussion Select an answer first