Announcing SOCI-indexed images for faster SageMaker AI jobs

This article announces SOCI-indexed images for Amazon SageMaker AI, which lazy-load container images to eliminate download overhead and accelerate job startup times.

• SOCI indexing fetches only files needed to start containers, streaming the rest in background while code runs
• Reduces container startup from minutes to seconds across all image sizes (8x–46x improvement)
• Pre-indexed AWS Deep Learning Containers available with -soci suffix; no API changes required
• Create custom SOCI indexes using nerdctl or Finch on Linux hosts with containerd and SOCI snapshotter
• Training performance unaffected after first epoch; net job savings of 1–6 minutes per job
• Automate index creation in CI/CD pipelines using CFN ECR SOCI Index Builder
• Available in private Amazon ECR repositories across all AWS Regions

SOCI indexing accelerates SageMaker AI job iteration by cutting container download time from minutes to seconds, reducing costs and infrastructure wait time without requiring changes to existing job definitions.