Scale-Up Networking

Overview

Scale-up networking is the load-store interconnect fabric within a rack or node that connects GPUs for intra-rack communication. Approximately 80% of GPU bandwidth flows over scale-up links, making it the dominant networking tier by volume for AI training workloads. It handles the tight coupling required between GPUs within a single training step, operating at the memory-level fabric layer rather than the packet network layer.

In AMD's three-tier AI networking model (scale-up, scale-out, scale-across), scale-up handles all intra-rack GPU-to-GPU communication, scale-out connects multiple racks within a data center, and scale-across links GPU clusters across multiple sites. AMD's Helios rack (announced for 2026) includes an integrated scale-up fabric built inside the rack.