What Is SuperNIC?_
A SuperNIC is a specialized network interface card that handles scale-out traffic (e.g., InfiniBand or Ethernet) between GPU servers, distinct from the NVLink-based GPU-to-GPU interconnect within a node or via NVSwitch across nodes. It typically connects via fiber optic cables and MPO connectors to top-of-rack switches. SuperNICs are essential for multi-node training jobs, providing the high bandwidth and low latency needed for distributed computing.
Technical Details
SuperNICs are installed in GPU servers and use fiber optic transceivers (e.g., QSFP or OSFP) to link to fabric switches. They operate at speeds such as 400 Gbps or 800 Gbps per port, depending on the generation. Unlike NVLink, which is typically a copper-based interconnect for direct GPU communication within a node or across a small number of nodes via NVSwitch, SuperNICs handle inter-rack and inter-cluster traffic. They require careful cable management and alignment of fiber pairs to maintain signal integrity.
How Leviathan Systems Works with SuperNIC
In our field deployments, we install SuperNICs in each GPU server and route fiber optic cables from them to the fabric switches in the top-of-rack or leaf-spine network. We verify link status and ensure proper cable polarity during commissioning.
Related service
Network Testing & Commissioning →