Scale-Across Networking

Overview

Scale-across networking connects GPU clusters across multiple buildings or geographic sites, driven by the practical impossibility of fitting one million GPUs in a single data center due to power, cooling, space, or availability constraints. Unlike scale-out (intra-DC), scale-across must handle asymmetric link speeds — some inter-site links may be 100 Gbps while others are 400 Gbps — requiring the NIC transport protocol to handle non-uniform bandwidth and ensure fairness across concurrent training applications sharing the constrained WAN paths.

AMD identified scale-across as the third tier of their AI networking model (alongside scale-up and scale-out), and the Pensando Vulcano 800G NIC (2026) is designed to handle all three tiers. The transport layer must account for WAN variability, asymmetric paths, and fair bandwidth allocation across multiple co-running applications.