Software-Defined Networking

Overview

Software-Defined Networking (SDN) decouples the network control plane (deciding how traffic flows) from the data plane (forwarding traffic), enabling programmatic control of network behavior from a central controller or distributed orchestration layer. In AI data centers, SDN is implemented primarily through open network operating systems (NOS):

SONiC (Microsoft-originated, widely adopted in hyperscale): Open-source, SAI-based, with broad ecosystem support across multiple ASIC vendors.
FBOSS (Meta's NOS): SAI-based, customized specifically for Meta's hyperscale deployment requirements; used in Meta's DSF (Distributed Switch Fabric) and NSF (Non-Scheduled Fabric) architectures.
Proprietary NOS: Legacy options from traditional networking vendors.

For AI cluster networking, SDN orchestration must extend beyond switching hardware to encompass GPUs, servers, and DLC (direct liquid cooling) management environments. When a training job stalls due to a network fault, rapid detection and remediation requires a unified orchestration view spanning compute, networking, and cooling. This is a key gap in current SDN deployments.

The 2025–2030 roadmap projects SDN transitioning to AI-based predictive orchestration starting around 2026 — using machine learning to anticipate network faults, intelligently triage events (distinguishing actual faults from expected workload behavior), and automate traffic optimization at scale.