Skip to content

Chapter 12: Scale-Up Networking -- NVLink Switch System · Part 2 of 8

Act 1 -- Two fabrics, two jobs

Inside a DGX H100 node, 4 third-generation NVSwitch chips provide 13.6 Tb/s of all-to-all bandwidth across 8 H100 GPUs. Every GPU can reach every other GPU in a single hop at 900 GB/s per GPU. This is the scale-up fabric -- NVLink, intra-node, designed for the tightest possible GPU-to-GPU communication.

Outside the DGX node, 8 ConnectX-7 HCAs connect to the external IB or RoCEv2 fabric at 400 Gbps each. A fully loaded DGX node contributes 3.2 Tbps of compute bandwidth to the external fabric. This is the scale-out fabric -- designed for communication between nodes.

NVLink — Intra-node vs Scale-out

Toggle between the two fabric tiers. The bandwidth cliff at the node boundary is the problem NVLink Switch solves.

DGX H100 Node (single chassis)NVSwitch 1NVSwitch 2NVSwitch 3NVSwitch 4GPU 0GPU 1GPU 2GPU 3GPU 4GPU 5GPU 6GPU 713.6 Tb/s all-to-all · 900 GB/s per GPU · 1 hop

All 8 GPUs are one NVSwitch hop away from each other. The 4 NVSwitch chips provide full all-to-all connectivity at 13.6 Tb/s aggregate. Every AllReduce that stays within this node runs at full NVLink bandwidth.

The two fabrics have completely different bandwidths, latencies, and addressing models:

Scale-up (NVLink intra-node)Scale-out (IB / RoCEv2)
Bandwidth per GPU900 GB/s400 Gbps (50 GB/s)
Bandwidth ratio18x fasterbaseline
Latency~1 us1-2 us
AddressingNVLink address spaceGID / GUID
TopologyFull-mesh via NVSwitchFat-tree
All-to-all (8 GPUs)13.6 Tb/s--

For a single DGX node, this asymmetry does not matter -- all GPU-to-GPU traffic stays on NVLink. The problem emerges the moment a job needs more than 8 GPUs. Any AllReduce that crosses the node boundary immediately drops from 900 GB/s NVLink bandwidth to 50 GB/s IB bandwidth per GPU -- an 18x cliff.

This cliff has a name: the scale-up boundary. Every model parallelism decision, every tensor partition, every pipeline stage assignment is affected by where this boundary falls.