Accelerating AI model production at Hexagon with Amazon SageMaker HyperPod

This article describes how Hexagon, a measurement technology leader, accelerated AI model production using Amazon SageMaker HyperPod for training specialized point cloud segmentation models.

• Hexagon develops domain-specific AI models for point cloud processing in aerospace, manufacturing, and geospatial applications
• SageMaker HyperPod provides resilient architecture with automated node failure recovery and job resumption
• Scalable infrastructure uses Elastic Fabric Adapter for optimal multi-node GPU communication
• Solution integrates Amazon S3, FSx for Lustre, and MLflow for data pipeline and experiment tracking
• Training time reduced 95% from 80 days on-premises to 4 days on AWS with 48 GPUs
• Larger batch sizes enabled better model accuracy and performance
• Flexible training plans provide predictable pricing and dedicated GPU capacity

Hexagon achieved significant efficiency gains through SageMaker HyperPod's scalable, resilient infrastructure, enabling faster AI model development and deployment.