Three-Computer Architecture

Overview

NVIDIA's three-computer architecture divides physical AI system development into three distinct compute tiers: (1) AI training on DGX supercomputers using foundation models, (2) simulation and digital twin on OVX servers running Omniverse, Isaac, and Cosmos for synthetic data generation and model validation, and (3) on-device edge inference via IGX and AGX running Holoscan for real-time sensor integration. The three tiers form a continuous closed-loop: models trained in the data center are validated in simulation before deployment to physical robots and devices, and feedback from edge deployments feeds back to training to refine models over time.

Inventec has implemented this architecture in its Physical AI Three-Computer DevKit — a scalable Kubernetes-orchestrated rack cluster with nodes in multiples of three, enabling growth from a two-rack MVP to a full scalability layout. The cluster directly instantiates the three-computer concept with Inventec Edge IGX Orin nodes (inference), Inventec P90G6 OVX servers (simulation), and DGX-tier training accessible via cloud or on-premises.