Describe Solutions to Maximize Performance in AI Workloads CERTIFIED-AI-EXPERT Practice Questions (Page 3)

Questions 11–15

1. 11application · medium

   A team trains a model on a dataset stored in an object store (e.g., S3). Each epoch, the data loader downloads images from the object store over the network, and GPU utilization is low because of network latency. The dataset is 1 TB and is accessed repeatedly each epoch. Which strategy would most effectively reduce the data-access latency?

   • AUse a slower network connection to reduce congestion.
   • BCache the dataset on local NVMe drives on each worker node and prefetch the next batch.
   • CIncrease the number of epochs to reuse the downloaded data more.
   • DReduce the image resolution to shrink the dataset size.
   Reveal Solution Discussion Select an answer first
2. 12application · medium

   A team is training a deep learning model and notices that training is slower than expected. They want to identify whether the bottleneck is in data loading, GPU compute, or network communication. Which tool or approach would best help them pinpoint the bottleneck?

   • AReduce the model size to see if training speeds up.
   • BCheck the GPU temperature to see if it is overheating.
   • CUse a profiler like NVIDIA Nsight Systems or PyTorch Profiler to trace GPU utilization, data-loader time, and communication time.
   • DIncrease the batch size and see if training speeds up.
   Reveal Solution Discussion Select an answer first
4. 13foundation · medium

   Which of the following is a common symptom of a network bottleneck in a distributed AI training workload?

   • AThe CPU is idle while the GPU is busy.
   • BThe training throughput scales poorly as more nodes are added.
   • CGPU utilization is high, but the loss function is not decreasing.
   • DThe storage system reports high latency.
   Reveal Solution Discussion Select an answer first
5. 14foundation · medium

   Which trade-off is most directly associated with using mixed precision training?

   • AImproved performance at the potential cost of numerical precision.
   • BHigher accuracy at the expense of training time.
   • CSimpler implementation at the expense of scalability.
   • DReduced cost at the expense of model size.
   Reveal Solution Discussion Select an answer first
6. 15foundation · medium

   Which technique is most effective for reducing data ingestion latency in an AI pipeline?

   • APrefetching data while the GPU is computing.
   • BUsing a larger batch size for training.
   • CReducing the model's learning rate.
   • DIncreasing the number of training epochs.
   Reveal Solution Discussion Select an answer first