This article announces the preview of Amazon SageMaker Profiler, a tool for tracking and visualizing hardware performance during deep learning model training on AWS.

• Tracks CPU and GPU activities including utilization, kernel runs, memory operations, and data transfers
• Provides Python modules for annotating PyTorch and TensorFlow training scripts
• Offers UI dashboard with visualizations of GPU/CPU active time, utilization trends, and kernel performance
• Includes timeline interface showing detailed kernel launches and runs at operation level
• Supports PyTorch 2.0.0, 1.13.1 and TensorFlow 2.12.0, 2.11.1 on specific GPU instance types
• Available in US East, US West, and Europe regions
• Generates up to 10x less profiling data than open-source alternatives

SageMaker Profiler helps ML practitioners optimize resource utilization and reduce training costs by identifying performance bottlenecks in large-scale distributed training jobs.