A team is training a large language model on a GPU cluster. They notice that GPU utilization is only 40% on average, while CPU utilization is 90%. Training is taking longer than expected. What is the most likely bottleneck and what should they do?

The model is too large for the GPU memory; they should reduce the batch size.

The GPUs are too slow; they should request a newer GPU model to speed up training.

The GPUs are underutilized because the data loading pipeline is the bottleneck; they should optimize data loading, such as using prefetching and parallel data loading.

The training is progressing normally; they should just wait longer.

Which practice is part of training data versioning?

Using the same data for all experiments

Recording a unique identifier for each data version

A team is running a large hyperparameter tuning campaign. They are using an experiment tracking tool, but they notice that some runs are missing metrics because the training script crashed before logging. What should they do to improve the reliability of experiment tracking?

Log metrics in real-time during training, not just at the end, and also log the configuration before training starts.

Only log metrics at the end of training to avoid overhead.

Manually record the metrics after each run from the console output.

Use a different experiment tracking tool that is more reliable.

What is the purpose of monitoring training progress using logs and metrics?

To detect issues like divergence or overfitting

To automatically tune hyperparameters

To replace the need for validation data

To reduce the size of the training dataset

PMICertified Professional in Managing AI (PMI-CPMAI)

Domain 4Objective 3

Manage AI/ML Model Training PMI-CPMAI Practice Questions (Page 6)

Part of the Manage AI Model Development and Evaluation domain, which accounts for 16% of the PMI-CPMAI exam. PMI does not publish an official question count, but from its 160-minute exam (~65–105 total, ~10–17 in this domain), expect 2–3 from this objective — we provide 36 practice questions to prepare you well beyond it. (estimate)

36questions here

8free pages

10concepts

16%of the exam

Questions 26–30

26application · medium
A team is training a large language model on a GPU cluster. They notice that GPU utilization is only 40% on average, while CPU utilization is 90%. Training is taking longer than expected. What is the most likely bottleneck and what should they do?
Select an answer first
27foundation · easy
Which practice is part of training data versioning?
Select an answer first
28expert · hard
A team is running a large hyperparameter tuning campaign. They are using an experiment tracking tool, but they notice that some runs are missing metrics because the training script crashed before logging. What should they do to improve the reliability of experiment tracking?
Select an answer first
29foundation · easy
What is the purpose of monitoring training progress using logs and metrics?
Select an answer first
30foundation · easy
Which tool is commonly used to track computational resource utilization during training?
Select an answer first

Finished these 5 questions?

Review the revealed explanations, or continue through the curriculum.

Next 5 questions →Back to PMI-CPMAI

Free Basic Practice is a study aid with revealable answers — not a scored exam. Examers.io is independent and not affiliated with or endorsed by PMI. “PMI-CPMAI” is a trademark of its owner, used for identification only.