An ML engineer trains a neural network in a SageMaker Studio notebook using an estimator. The Python training script uses Distributed Data Parallel (DDP) on a single instance that has multiple GPUs, but the GPUs are underutilized. The engineer needs to find where in the training script to optimize resource usage. Which approach will identify that point?
Choose an answer
Tap an option to check your answer.
Correct answer: Add SageMaker Profiler annotations in the training script, run it, and generate a profiler report from the results..
Why this is the answer
SageMaker Profiler is designed to help identify performance bottlenecks within training jobs, including GPU underutilization. By adding annotations to the training script, the profiler can collect detailed metrics on operations, memory usage, and GPU activity, generating a report that pinpoints the exact code sections causing underutilization. CloudWatch metrics provide high-level GPU utilization but lack the granular detail needed to identify specific code bottlenecks within the training script. CloudTrail logs API calls and user activity, not detailed resource utilization or performance metrics for training jobs. SageMaker Model Monitor is used for monitoring model quality and data drift in production, not for optimizing training job performance.
Pass your exam — without the endless answer hunt
Get every verified question and explanation for this exam in one place, and save hours of prep. 1,000+ certifications · 20+ languages · free to start.
Pass your exam faster → No card needed