What Is Top-of-Rack Switch?_
In GPU data centers, a ToR switch sits at the top of each rack, providing the first layer of network connectivity for the servers within that rack. It handles both intra-rack traffic (between servers in the same rack) and uplinks to higher-level spine switches for inter-rack communication. ToR switches are typically 1U or 2U form factors and use high-density ports like QSFP or OSFP for fiber connections to the spine.
Technical Details
ToR switches in GPU clusters often support 100 Gbps or 400 Gbps per port, with uplinks to the spine network using single-mode or multimode fiber. They are separate from the NVLink interconnect fabric (e.g., NVSwitch), which is a copper-based GPU-to-GPU network internal to the rack. The switch's latency and throughput must match the fabric's requirements, as specified by the OEM for the given network standard (e.g., InfiniBand or Ethernet).
How Leviathan Systems Works with Top-of-Rack Switch
During rack assembly, we mount the ToR switch at the top of the rack, run structured cabling from each GPU server's NIC to the switch's front ports, and connect uplinks to the spine network. We also verify power and management connections before commissioning.
Related service
Network Testing & Commissioning →