AI Network Designer Design a network →

Server specifications

The network face of each AI platform: how many rails it presents, at what speed, and what it needs from the storage and management fabrics.

The compute rail count is the number that drives the fabric. In a rail-optimized design it sets how many independent leaf groups exist, and each is sized on its own before being multiplied back up. Storage and in-band are separate fabrics with their own switches and optics.

Rack-scale platforms are quoted per rack, not per node. For those, the figures below cover the scale-out network only. The NVLink domain inside the rack is a separate fabric that ships with the rack and is not something you buy switches for.

NVIDIA

PlatformGPUsCompute railsStorageIn-bandTypical kW
DGX H100 8 8 × 400G 2 × 400G 2 × 100G 10.2
DGX H200 8 8 × 400G 2 × 400G 2 × 100G 10.2
DGX B200 8 8 × 400G 2 × 400G 2 × 100G 14.3
DGX B300 8 8 × 800G 2 × 400G 2 × 100G 13
DGX B300 — Spectrum-X 8 16 × 400G 2 × 400G 2 × 100G 13
GB200 NVL72 72 72 × 400G 36 × 400G 36 × 100G 120
GB300 NVL72 72 72 × 800G 18 × 400G 18 × 100G 120
GB300 NVL72 — Spectrum-X 72 144 × 400G 18 × 400G 18 × 100G 120

Cisco

PlatformGPUsCompute railsStorageIn-bandTypical kW
Cisco UCS C885A M8 8 8 × 400G 2 × 400G 2 × 100G 10.2
Cisco UCS C880A M8 8 8 × 800G 2 × 400G 2 × 100G 13
Cisco UCS C845A M8 8 4 × 400G 1 × 200G 1 × 100G 6.5

Supermicro

PlatformGPUsCompute railsStorageIn-bandTypical kW
Supermicro SYS-821GE-TNHR 8 8 × 400G 2 × 400G 2 × 100G 10.2
Supermicro SYS-421GE-TNHRT 8 8 × 400G 2 × 400G 2 × 100G 10.2
Supermicro AS-8125GS-TNHR 8 8 × 400G 2 × 400G 2 × 100G 10.2
Supermicro SYS-421GE-TNHR2-LCC 8 8 × 400G 2 × 400G 2 × 100G 10.2
Supermicro SYS-421GE-NBRT-LCC 8 8 × 400G 2 × 400G 2 × 100G 14
Supermicro SYS-A21GE-NBRT 8 8 × 400G 2 × 400G 2 × 100G 14
Supermicro HGX B300 8 8 × 800G 2 × 400G 2 × 100G 13
Supermicro GB300 NVL72 72 72 × 800G 18 × 400G 18 × 100G 120

Dell

PlatformGPUsCompute railsStorageIn-bandTypical kW
Dell PowerEdge XE9680 8 8 × 400G 2 × 400G 2 × 100G 8
Dell PowerEdge XE9640 4 4 × 400G 1 × 200G 1 × 100G 5.5

HPE

PlatformGPUsCompute railsStorageIn-bandTypical kW
HPE Cray XD670 8 8 × 400G 2 × 400G 2 × 100G 10.2
HPE ProLiant Compute XD685 8 8 × 400G 2 × 400G 2 × 100G 14

Lenovo

PlatformGPUsCompute railsStorageIn-bandTypical kW
Lenovo ThinkSystem SR675 V3 8 8 × 400G 2 × 400G 2 × 100G 10.2

Generic

PlatformGPUsCompute railsStorageIn-bandTypical kW
Generic HGX 8-GPU 8 8 × 400G 2 × 200G 1 × 100G 10
Generic 4-GPU 4 4 × 400G 1 × 200G 1 × 25G 5

Related