AI Network Designer Design a network →

GB200 and GB300 NVL72 networking

Three networks, not one. Only two of them are a purchasing decision, and the GB200 and GB300 trays are not the same hardware.

The common misunderstanding about NVL72 is that NVLink replaces the cluster network. It does not. A rack carries three distinct networks and they do different jobs.

NetworkScopePer rackDo you buy switches and optics?
NVLink Inside the rack, GPU to GPU 18 links per GPU to the in-rack NVSwitches, over a copper backplane, through 9 switch trays No. Ships with the rack
ConnectX (east/west) Between racks, GPU to GPU 72 adapters, a 1:1 GPU-to-NIC ratio Yes. This is the compute fabric
BlueField-3 (north/south) Storage, provisioning, management 18 or 36 links depending on generation Yes. Storage and in-band fabrics

NVLink is per-GPU and it is what makes 72 GPUs behave as one accelerator. It is also entirely internal: copper backplane, no optics, no cable schedule, no line in the bill of materials. You do not size it. You do size the other two.

GB200 and GB300 trays are different hardware

Both racks are 18 compute trays of 4 GPUs. The trays themselves differ in three ways that change the parts list.

GB200 trayGB300 tray
East/west 4× single-port 400G ConnectX-7 4× dual-port 800G ConnectX-8
Rack network total 28.8 Tb/s 57.6 Tb/s
GPU to NIC path GPU → Grace over C2C → NIC over PCIe NIC-direct; the GPU reaches the ConnectX-8 itself
North/south 2× dual-port BlueField-3, management ports bonded 1× dual-port BlueField-3 B3240, no bond
North/south links per rack 36 storage + 36 management 18 storage + 18 management

NVIDIA's own interface enumeration settles the DPU question. A GB200 tray presents M1 and M2 at PCI addresses enP6 and enP22, bonded as bond0 in LACP mode 4, plus S1 and S2 on storagenet. A GB300 tray presents only M1 and S1, both on enP22, the same address as the GB200's second card. The documentation states plainly that there is no bond configuration for GB300 trays, because there is only one card to bond.

A bond is redundancy, not consolidation. Two bonded management ports still land on two switch ports and still need two cables and two optics. Count a bond as one link and you halve the in-band fabric.

Ethernet changes the link count, not the bandwidth

ConnectX-8 runs one 800G port in InfiniBand mode and two 400G ports in Ethernet mode. A Spectrum-X build of a GB300 NVL72 therefore terminates 144 links at 400G instead of 72 at 800G. The rack still moves 57.6 Tb/s. It just presents twice as many terminations, and terminations are what switch, optic and cable counts are made of.

Rack granularity and power

The indivisible unit is a whole 72-GPU rack, so GPU counts round to multiples of 72 instead of 8. An NVL72 rack draws roughly 120 kW continuous at 132 kW nominal TDP, with peaks near 155 kW that the electrical design has to survive even though that is not the steady-state figure. It is liquid-cooled and needs a completely different power design from air-cooled DGX.

Related