The H100 CNX is a product that only makes sense once you understand both the GPU and the NIC sides of this course. It co-packages an H100 SXM GPU and a ConnectX-7 HCA on a single module, connected by PCIe Gen 5 -- not by NVLink.

The design target is InfiniBand-only smart endpoint deployments where the goal is to minimise the PCIe hops between GPU and NIC. In a standard DGX H100, the path from GPU HBM to NIC is: GPU HBM -> NVLink -> CPU PCIe root complex -> PCIe -> ConnectX-7. In an H100 CNX, the GPU and the ConnectX-7 are on the same module, and the NIC connects directly to the GPU's PCIe interface without traversing the CPU PCIe root complex. This reduces the RDMA write path from 4 hops to 2 hops.

H100 CNX is an InfiniBand-only product -- it only supports IB mode, not Ethernet. It is targeted at very large-scale HPC deployments where the NIC-GPU path efficiency matters more than Ethernet/RoCE compatibility.

For most practitioners designing DGX-based clusters, H100 CNX is not the relevant product -- the standard DGX H100 with 8 single-port compute ConnectX-7s plus 2 dual-port Slot1/Slot2 ConnectX-7s for storage/management is the correct reference architecture. H100 CNX is mentioned here because it appears in NVIDIA's product literature and represents a specific architectural bet (PCIe over NVLink-C2C for GPU-NIC coupling, IB-only) that differs from the DGX approach.