CONTACT

How Many Optical Transceivers Does an NVIDIA H100 Cluster Really Need?

A practical breakdown of transceiver requirements in NVIDIA DGX H100 SuperPOD architectures using 400G and 800G optics.

June 22, 2024

Transceivers Needed in NVIDIA H100 Server Cluster

Are you familiar with the network architecture of the NVIDIA DGX H100 SuperPOD? Have you had a chance to consider the number of optical transceivers that might be needed? Let’s explore this together, using the NVIDIA 4 SU solution as a guide to estimate the number of optical transceivers.
In the H100+ConnectX-7+QM9700 two-layer network architecture, there are 128 servers, 32 Leaf switches, and 16 Spine switches.

 

1536 Pieces 800G Optical Transceivers Are Required：

Each Leaf switch has 16 interfaces(800G) for upward transmission and 16 interfaces(800G) for downstream transmission. Therefore, a total of 32*16+32*16=1024 pieces optical transceivers(800G) are required on 32 Leaf switches. Each Spine switch has 32 interfaces(800G), requiring 512 pieces optical transceivers(800G) for the 16 Spine switches.

 

1024 Pieces 400G Optical Transceivers Are Required:

Each Leaf has 16 downward transmission interfaces, and each interface connects downward to the server at a unidirectional rate of 2*400G. Therefore, a server connected to 32 Leaf switches requires 32*16*2=1024 pieces optical transceivers(400G).

Transceivers Needed in NVIDIA H100 Server Cluster

Plan with confidence

If you’re designing or scaling an NVIDIA-based AI cluster, ATOP can help validate optical requirements, configurations, and sourcing options.
Contact ATOP
EDIT
arrow-right