What Is Leaf Switch?_
In a leaf-spine architecture, leaf switches provide the access layer for GPU servers, handling east-west traffic between nodes and north-south traffic to the spine layer. They typically host high-density ports for 400/800 GbE or NDR InfiniBand, and are deployed in pairs for redundancy. Each leaf switch aggregates multiple GPU servers and forwards traffic to the spine switches for cross-rack communication.
Technical Details
Leaf switches in GPU clusters use QSFP-DD or OSFP cages for high-speed optics, and are connected to GPU servers via copper DAC cables or active optical cables (AOC) depending on distance. They must support lossless RoCEv2 or InfiniBand transport for RDMA workloads, requiring PFC and ECN configuration. In liquid-cooled racks, leaf switches are often placed in a separate air-cooled zone or at the top of the rack with dedicated airflow. The leaf-to-spine oversubscription ratio is set per the fabric design, typically 1:1 for non-blocking performance.
How Leviathan Systems Works with Leaf Switch
Our field crew installs leaf switches in each GPU rack, cabling them to GPU server NICs and to spine switches in the central patch panel. We verify link bring-up and RDMA connectivity during commissioning, often using a pre-configured leaf switch image from the customer's network team.
Related service
Network Testing & Commissioning →