What Is 400GbE?_
400 Gigabit Ethernet (400GbE) is the current mainstream Ethernet speed for data center switch-to-server connections in GPU clusters. 400GbE uses QSFP-DD or OSFP transceivers with PAM4 signaling. Connections are made via DAC (intra-rack), AOC (inter-rack short), or transceivers with fiber patch cables (longer runs). 800GbE is emerging for new deployments.
Technical Details
400GbE uses IEEE 802.3bs and related standards, employing PAM4 (4-level pulse amplitude modulation) signaling to achieve higher data rates per lane. Common 400GbE transceiver options include 400GBASE-SR8 (multimode, 8 lanes, up to 100m), 400GBASE-DR4 (single-mode, 4 lanes, up to 500m), and 400GBASE-FR4 (single-mode, 4 lanes, up to 2km). The transition from 100GbE to 400GbE quadruples per-port bandwidth, reducing the number of connections needed or enabling higher aggregate bandwidth. For GPU clusters, 400GbE has become the standard Ethernet speed for server-to-leaf and leaf-to-spine connections, with multiple ports per GPU node providing the aggregate bandwidth needed for distributed training.
How Leviathan Systems Works with 400GbE
Leviathan Systems deploys 400GbE infrastructure in GPU data centers, including cable selection (DAC, AOC, fiber), transceiver configuration, and testing to verify link integrity at 400G speeds.