GB200 interconnects: separate the fabrics before counting fiber.

What NVL72 identifies.

In NVIDIA’s documented DGX GB200 NVL72 configuration, 18 compute trays provide four GPUs each, connected through nine NVLink switch trays. The resulting scale-up domain contains 72 GPUs. The rack’s NVLink connection uses a passive copper cable-cartridge backplane, while external network interfaces serve different fabrics. DGX GB rack hardware ↗

That distinction is the first step in an optics plan. Counting every GPU-to-NVLink-switch connection as an optical port would create a large and incorrect transceiver requirement. The presence of liquid cooling also says nothing by itself about the connector, optical lane arrangement or cooling interface of an external module.

Draw the connection domains separately.

The cited DGX GB200 SuperPOD architecture describes NVLink, compute, storage, in-band management and out-of-band management as five logical fabrics carried over four physical fabrics. Logical separation does not necessarily mean five separate cable plants. Copy the actual platform design into your connection schedule. SuperPOD network fabrics ↗

Normalize bandwidth units and direction.

NVIDIA specifies NVLink 5 at 1.8 TB/s bidirectional per GPU in this system context. A 400Gb/s network-port label uses bits per second and conventionally states the rate in one direction. Comparing 1.8 to 400 without converting units and direction produces meaningless ratios. DGX GB networking reference ↗

For an illustrative unit conversion, 1.8 TB/s is 14.4 Tb/s when both directions remain summed. A symmetric directional split would be 7.2 Tb/s each way. Neither figure is an optical-port count or a guarantee of application throughput. Similarly, 400Gb/s is 50 GB/s before protocol overhead.

Keep GPU injection bandwidth, switch capacity, fabric bisection bandwidth and workload throughput on separate lines in the design. Label whether each quantity is per GPU, per NIC, per tray, per rack or per fabric.

Worked count: 72 links need not mean 144 modules.

The documented DGX GB200 tray has four ConnectX-7 400G compute NICs. Assume 18 trays, one 400G optical connection per NIC, and supported twin-port 2×400G optical modules at the switches. Also assume each NIC uses one single-port module and one straight fiber assembly. Tray configuration · Twin-port connection structure

The example excludes switch-to-switch connections, storage, management, intermediate patch cords and spares. A different optical presentation or port grouping changes the physical bill. Count usable switch ports before applying this arithmetic; do not infer capacity solely from the number of visible cages.

Carry the map from rack to row.

Assign each NIC to its intended rail or plane before assigning patch-panel positions. Record the switch port and sub-port at the far end, then define the actual fiber route and service slack. This makes a cross-connection discoverable without relying on a technician to infer the topology from cable colors.

Reserve a separate schedule for fabric uplinks. It should show link rate, number of parallel links, peer location and failure grouping. A rack’s endpoint count does not determine leaf-to-spine quantities until the topology and oversubscription policy are fixed.

Preserve route lengths and connector access around manifolds and service areas. The mechanical layout must permit both thermal service and optical inspection without making one dependent on dismantling the other.

Validate each fabric on its own terms.

NVLink topology health, external NIC link health and storage reachability are different acceptance results. Verify each, then run a workload that crosses the intended domain boundary. Save the rack/port inventory, firmware context, connection drawing and measured observations together.

For a GB300 design, re-read the corresponding platform specification instead of substituting “800G” into this GB200 example. The NIC generation, optical lane mode and network architecture must be reconciled together.

Sources and applicability

Original WBN architecture and counting guidance. NVIDIA references checked 2 November 2026. The numeric example is derived from stated assumptions and excludes fabric uplinks, services and spares; it is not a complete deployment bill of materials.

1. NVIDIA DGX GB Rack Scale Systems — Hardware

   Rack, compute-tray, copper backplane and GB200/GB300 NIC configuration. Online guide checked 2 November 2026.

2. NVIDIA DGX GB200 SuperPOD — Network Fabrics

   Reference-architecture logical versus physical fabrics and scale-out topology.

3. NVIDIA DGX GB Rack Scale Systems — Networking

   NVLink bandwidth convention and distinct platform networks.

4. NVIDIA LinkX — Transceivers and Fiber Details

   100G-PAM4 twin-port/single-port structure used in the explicit counting assumptions.

Continue the engineering record

• DGX B300 versus GB300 NVL72
• ConnectX-7/8 optics and OSFP IHS/RHS
• Duplex, parallel and breakout fiber paths