What Is NVSwitch?_
NVSwitch is NVIDIA's custom silicon that connects multiple GPUs via NVLink within a server node or rack. Inside an HGX H100 baseboard, NVSwitch provides all-to-all connectivity between 8 GPUs. In the NVL72 rack systems, multiple NVSwitch chips create a unified NVLink fabric connecting all 72 GPUs. NVSwitch connections are internal and do not require external cabling by the deployment team.
Technical Details
NVSwitch is a purpose-built ASIC that serves as a crossbar switch for NVLink connections. In the HGX H100, four NVSwitch chips connect 8 GPUs with all-to-all NVLink 4.0 connectivity at 900 GB/s per GPU. In NVL72 systems, NVSwitch chips are distributed across the rack's compute trays and switch trays, creating a unified NVLink 5.0 fabric with 130 TB/s total bandwidth across all 72 GPUs. This means any GPU can communicate with any other GPU at full NVLink speed without traversing the external network. NVSwitch connections are entirely internal to the baseboard or rack, so deployment teams do not handle NVSwitch cabling. However, understanding NVSwitch topology is important for troubleshooting and performance validation.
How Leviathan Systems Works with NVSwitch
Leviathan Systems validates NVSwitch functionality during commissioning by verifying that all GPU-to-GPU NVLink paths are operational, ensuring the internal interconnect fabric is performing as specified.