Hexagon, a global leader in measurement technologies, has revolutionized its AI development pipeline by leveraging Amazon SageMaker HyperPod. The company specializes in processing point clouds—digital maps made of millions of individual 3D data points that represent the external surfaces of objects or landscapes. These complex data sets are foundational for creating digital twins of entire cities and managing heavy construction sites.

Previously, training the specialized models required to clean and segment this 3D data took 80 days using on-premises hardware. By migrating to a high-performance cluster managed by SageMaker HyperPod, Hexagon compressed this timeline to just four days, representing a staggering 95% improvement in efficiency. This shift allows the company to iterate on new AI use cases in days rather than months, significantly shortening their time-to-market for vital infrastructure solutions.

The new system utilizes a resilient architecture that automatically detects and replaces faulty computing components. This "self-healing" capability ensures that long-running training jobs resume from the most recent checkpoint without human intervention, preventing wasted effort during hardware failures. To handle the massive data flow, Hexagon uses a specialized file system that streams terabytes of training data directly to processors at high speeds.

This collaboration demonstrates how managed infrastructure removes the "heavy lifting" of cluster management. By automating health checks and optimizing communication between servers, Hexagon’s engineers can focus on refining model accuracy for aerospace, automotive, and geospatial sectors. The result is a more precise and faster understanding of the built environment across various global industries.

Hexagon, a global leader in measurement technologies, has revolutionized its AI development pipeline by leveraging Amazon SageMaker HyperPod. The company specializes in processing point clouds—digital maps made of millions of individual 3D data points that represent the external surfaces of objects or landscapes. These complex data sets are foundational for creating digital twins of entire cities and managing heavy construction sites.

Previously, training the specialized models required to clean and segment this 3D data took 80 days using on-premises hardware. By migrating to a high-performance cluster managed by SageMaker HyperPod, Hexagon compressed this timeline to just four days, representing a staggering 95% improvement in efficiency. This shift allows the company to iterate on new AI use cases in days rather than months, significantly shortening their time-to-market for vital infrastructure solutions.

The new system utilizes a resilient architecture that automatically detects and replaces faulty computing components. This "self-healing" capability ensures that long-running training jobs resume from the most recent checkpoint without human intervention, preventing wasted effort during hardware failures. To handle the massive data flow, Hexagon uses a specialized file system that streams terabytes of training data directly to processors at high speeds.

This collaboration demonstrates how managed infrastructure removes the "heavy lifting" of cluster management. By automating health checks and optimizing communication between servers, Hexagon’s engineers can focus on refining model accuracy for aerospace, automotive, and geospatial sectors. The result is a more precise and faster understanding of the built environment across various global industries.