AI Network Designer Design a network →

64 GPU AI cluster network: GB300 NVL72

You cannot buy 64 GPUs of NVL72: the rack is the unit, so this rounds up to 72. Those 72 links at 800G fit one Quantum-X800, and a lone rack is already one accelerator over NVLink anyway.

The build

PlatformNVIDIA GB300 NVL72, 72 × GB300 (Blackwell Ultra) per rack
Compute fabricinfiniband, 72 rails at 800G
Leaf1 × Q3400-RA (Quantum-X800)
Spinenot required at this scale
Corenot required at this scale
Racks1 compute + 1 network, 1 rack per rack, limited by fixed
Power120 kW compute + 9 kW network = 129 kW

How this fabric was sized.

72 endpoints at 800G fit a single Q3400-RA (Quantum-X800) with 72 ports spare. No spine tier is needed and rail optimization buys nothing here: every GPU is already one hop from every other. A rail-optimized arrangement at this size would cost one leaf per rail plus spines to join them, for no gain. Note this is a single point of failure for the compute fabric; a second switch for redundancy is a resilience decision rather than a capacity one.

Asked for 64, built 72. An NVL72 rack is indivisible, so GPU counts round to multiples of 72. 1 racks is the smallest build that meets the target.

Fabrics

Compute

72 endpoints across 72 rails, 1 tier, realised oversubscription 1:1 (non-blocking).

TierSwitchCountDown/swUp/swPorts used
leaf NVIDIA Q3400-RA (Quantum-X800) 1 144 0 72 / 144

Storage

50 endpoints across 1 rail, 2 tiers, realised oversubscription 1.91:1.

TierSwitchCountDown/swUp/swPorts used
leaf NVIDIA QM9700 (Quantum-2) 2 42 22 94 / 128
spine NVIDIA QM9790 (Quantum-2, externally managed) 1 64 0 44 / 64

In-band management

18 endpoints across 1 rail, 2 tiers, realised oversubscription 3.67:1.

TierSwitchCountDown/swUp/swPorts used
leaf Cisco Nexus 93600CD-GX 1 22 6 24 / 28
spine Cisco Nexus 93600CD-GX 1 28 0 6 / 28

Out-of-band management

29 endpoints across 1 rail, 1 tier, realised oversubscription 1:1 (non-blocking).

TierSwitchCountDown/swUp/swPorts used
leaf Cisco Nexus 9348GC-FX3 1 48 4 29 / 48

Bill of materials

Every line carries the calculation that produced its quantity. Prices are deliberately absent: the catalog ships without list prices, because inventing them would be worse than leaving them out.

ItemQtyWhy this quantity
Servers
GB300 NVL72 (72 x GB300 (Blackwell Ultra)) 1 64 GPUs requested / 72 GPUs per node = 1 nodes (72 GPUs).
Network adapters
NVIDIA ConnectX-8 SuperNIC C8180 (XDR 800G)
900-9X81E-00EX-ST0
72 1 nodes x 72 compute NICs per node (one per rail).
Compute fabric switches
NVIDIA Q3400-RA (Quantum-X800)
920-9B36F-00RX-8S0
1 1 across Compute fabric (single switch) leaf.
NVIDIA QM9700 (Quantum-2)
MQM9700-NS2F
2 2 across Storage fabric leaf.
NVIDIA QM9790 (Quantum-2, externally managed)
MQM9790-NS2F
1 1 across Storage fabric spine.
Management switches
Cisco Nexus 9348GC-FX3
N9K-C9348GC-FX3
1 1 across Out-of-band management leaf.
Cisco Nexus 93600CD-GX
N9K-C93600CD-GX
2 2 across In-band management leaf, In-band management spine.
Optics — transceivers
800G OSFP 2xSR4 twin-port (NDR)
MMA4Z00-NS
46 44 x 400G Storage leaf to Storage spine at 30 m. 30 m run needs discrete optics over structured fiber. 800G OSFP 2xSR4 twin-port (NDR) at Storage leaf and 800G OSFP 2xSR4 twin-port (NDR) at Storage spine, over MPO-12 APC OM4 multi-mode trunk. The Storage leaf module is twin-port, so it terminates 2 links and only 0.5 module is needed per link.
Cables and assemblies
100G QSFP28 passive DAC 19 18 x 100G In-band NIC to In-band switch at 2 m. 2 m run is within the 2 m passive copper limit. Optics are integrated into the assembly, so no separate transceivers are needed.
800G OSFP passive DAC
MCP4Y10-N00A
75 72 x 800G Compute NIC to Leaf at 2 m. 2 m run is within the 2 m passive copper limit. Optics are integrated into the assembly, so no separate transceivers are needed.
800G OSFP to 2x400G OSFP splitter DAC
MCP7Y00-N001
26 18 x 400G Storage NIC to Storage leaf at 2 m. 2 m run is within the 2 m passive copper limit. Optics are integrated into the assembly, so no separate transceivers are needed; one cage breaks out to 2 links.
Cat6a patch lead 30 29 x 1G BMC / device management port to OOB switch at 2 m. 1G over structured copper, 2 m within the 100 m limit for Cat6a patch lead.
Structured fibre
MPO-12 APC OM4 multi-mode trunk 46 44 x 400G Storage leaf to Storage spine at 30 m. 30 m run needs discrete optics over structured fiber. 800G OSFP 2xSR4 twin-port (NDR) at Storage leaf and 800G OSFP 2xSR4 twin-port (NDR) at Storage spine, over MPO-12 APC OM4 multi-mode trunk. The Storage leaf module is twin-port, so it terminates 2 links and only 0.5 module is needed per link.
Racks
Rack 2 1 compute racks at 1 nodes each (fixed-limited) plus 1 network racks.
Power
Rack PDU 4 2 racks x 2 PDUs per rack.

Related