| Planning dimension | MGX | HGX |
|---|---|---|
| What it is | A modular reference architecture for OEM/ODM server designs. | An integrated multi-GPU platform built around an interconnected GPU baseboard. |
| Design focus | Flexible combinations of supported CPUs, GPUs, networking, and thermal designs. | High-bandwidth GPU communication for demanding AI and HPC workloads. |
| H200 example | H200 NVL PCIe systems; supported NVLink bridges connect 2 or 4 GPUs in a group. | H200 SXM systems; supported 8-GPU designs combine NVLink with NVSwitch. |
| What to confirm | Exact chassis, supported GPU SKU, bridge groups, host PCIe topology, and power/cooling. | Exact generation and baseboard, GPU count, fabric, OEM system, facility power, and cooling. |
MGX is not a GPU model or a promise of a particular NVLink topology. A full server configuration must identify the exact hardware and supported interconnect.