LISP-Based Network for AI Infrastructure

The Locator/Identifier Separation Protocol (LISP) control plane provides the mechanisms to support Scale-Up, Scale-Out and Scale- Across backend networks within AI infrastructure. This document describes how LISP enables a unified control plane architecture that accommodates different scaling technologies, offering flexibility in deployment. By leveraging lisp mechanisms including EID-to-RLOC- mapping/pxTR registrations, publication-subscription and VPN instance-based segmentation ([RFC9300][RFC9301][RFC9437][I-D.ietf- lisp-site-external-connectivity][I-D.ietf-lisp-vpn]), the architecture delivers scalable high bandwidth for large training workloads, ultra-low-latency collective operations, and a simplified control framework that seamlessly spans heterogeneous accelerator fabrics. This approach allows AI/ML applications (whether focused on training or inference) to run workloads efficiently with resiliency on a converged infrastructure, supporting diverse deployment scenarios using the same underlying network fabric.