Multi-Plane Networking

Overview

Multi-plane networking is a fabric architecture that splits GPU NIC bandwidth across multiple independent leaf-spine topologies (planes), rather than aggregating all GPU bandwidth into a single fat-tree. For a 100K GPU cluster, single-plane requires 5,000 switches while multi-plane requires only 3,000 — a 40% reduction in switch and optics capex. The architectural rationale is that splitting bandwidth across planes expands effective switch radix and reduces the number of Clos layers needed, lowering both cost and operational complexity.

Beyond cost, multi-plane provides fault isolation: a congestion event or switch failure in one plane is contained within that plane and does not cascade to other planes. AMD implements multi-plane on the Pollara 400 by splitting its 400 Gbps port into four 100 Gbps ports, with two levels of NIC load balancing: per-entropy RDMA flow splitting within a plane, and cross-plane distribution at the NIC level.