Describe Solutions to Maximize Performance in AI Workloads CERTIFIED-AI-EXPERT Practice Questions (Page 7)

Questions 31–35

1. 31application · hard

   A team wants to train a model with 50 billion parameters on a cluster of 8 GPUs, each with 40 GB of memory. The model does not fit on a single GPU. The team also has a very large dataset. Which parallelism strategy should they use to train this model?

   • AData parallelism only, since the dataset is large.
   • BUse a single GPU with gradient checkpointing to reduce memory.
   • CModel parallelism to split the model layers across GPUs, combined with data parallelism for the large dataset.
   • DPipeline parallelism only, without data parallelism.
   Reveal Solution Discussion Select an answer first
2. 32foundation · medium

   Which symptom most directly indicates a data I/O bottleneck in an AI training workload?

   • AGPU utilization is high while the CPU is idle.
   • BNetwork latency is high between nodes during distributed training.
   • CMemory usage is consistently near capacity.
   • DGPU utilization is low while the CPU is busy reading and preprocessing data.
   Reveal Solution Discussion Select an answer first
4. 33foundation · medium

   Which storage tier is most appropriate for active training data that requires high throughput and low latency?

   • AStandard HDD tier
   • BObject storage with infrequent access
   • CCold archive tier
   • DHigh-performance SSD tier
   Reveal Solution Discussion Select an answer first
5. 34expert · hard

   A team trains a model on a dataset of 100 TB stored on a parallel file system. Profiling shows that the data loader is the bottleneck, but the storage system has plenty of IOPS and bandwidth. The team notices that the file system uses a small stripe size (e.g., 64 KB), and the training files are large (e.g., 1 GB each). Which file system tuning would most likely improve data-loading throughput?

   • AReduce the stripe size to fit more stripes on each node.
   • BIncrease the stripe size or stripe count so large files are spread across more storage nodes.
   • CMove the dataset to a single large HDD to simplify access.
   • DDisable caching to reduce memory pressure.
   Reveal Solution Discussion Select an answer first
6. 35expert · hard

   A team trains a model and notices that GPU utilization is high (95%), but the loss is not decreasing as expected. Profiling shows that the forward pass is fast, but the backward pass is slow. The team suspects the model has a large number of parameters and the gradient computation is memory-bound. Which optimization would most directly speed up the backward pass?

   • AUse mixed precision training to reduce memory bandwidth and speed up gradient computation.
   • BSwitch to a smaller model to reduce gradient computation.
   • CReduce the number of GPUs to lower communication overhead.
   • DIncrease the batch size to improve GPU utilization.
   Reveal Solution Discussion Select an answer first