Announcing instance preference lists for Amazon SageMaker AI training jobs

This article announces instance preference lists for Amazon SageMaker AI training and processing jobs, enabling automatic GPU capacity provisioning across multiple instance types.

• Specify an ordered list of up to five acceptable instance types when creating training or processing jobs
• SageMaker automatically evaluates the list in priority order and launches on the first available type
• Eliminates manual retry loops and custom monitoring scripts for managing job submissions
• Integrates with Flexible Training Plans to prioritize reserved capacity before falling back to on-demand
• Supports per-preference instance counts for compute equivalency across different GPU architectures
• Jobs enter event-driven queue with automatic retry if no capacity available, bounded by MaxPendingTimeInSeconds
• Available for both training jobs and processing jobs via SageMaker Python SDK v3

Instance preference lists reduce GPU capacity management overhead, accelerate job starts, and improve resource utilization for AI model training and data processing workloads.