MSI CG480-S5063

Thirteen PCIe 5.0 x16 slots taking up to eight H200 NVL or RTX PRO 6000 Blackwell Server Edition cards. That is the argument for a PCIe system over an SXM one: a card can be added, changed or replaced, and the slots that are not carrying GPUs can carry network or storage adapters instead.

Twenty front hot-swap E1.S NVMe bays is an unusual amount of local storage for a 4U GPU server, and it is the reason to look at this machine specifically. A training job whose dataset fits on the node does not touch the network for reads at all, which removes the storage fabric from the critical path.

It is an NVIDIA-Certified System on the MGX architecture, so the software stack and the reference designs are the same ones the SXM machines use. The trade against an HGX platform is interconnect: PCIe cards do not have NVLink between them, so a model that has to span all eight GPUs will see a slower path than the same model on a B300 baseboard.