Operating NVIDIA OSMO-based Physical AI Workflows on Amazon EKS

Physical AI model development has a long lifecycle where data collection, simulation, policy training, and edge deployment repeat iteratively. To reliably run repetitive workflows, execution environments, artifact preservation, observability, security, and version compatibility must be managed consistently.

This architecture presents a reference model for safely and efficiently operating Physical AI workloads on Amazon EKS while keeping NVIDIA OSMO as an external dependency.

The key components are as follows:

• Foundational infrastructure: Amazon EKS is used as the control plane, and metadata and artifacts are safely preserved through AWS managed services such as Amazon RDS, S3, ECR, and Secrets Manager. Observability is secured through Amazon Managed Service for Prometheus and Grafana.
• GPU execution environment: Karpenter manages capacity for Amazon EC2 G family GPU instances, while NVIDIA GPU Operator and Elastic Fabric Adapter(EFA) provide resource exposure and network acceleration.
• Workflow execution: KAI Scheduler handles PodGroup CRDs and gang scheduling, while NVIDIA OSMO manages overall orchestration and metadata management.

Deployment is designed so that infrastructure and components can be installed step by step using Terraform, Helm, and kubectl. For security, all components operate within a private network, and data and volumes are managed with encryption.