What Is MGX Modular Architecture?_
MGX is a reference architecture for building dense GPU systems, typically used in NVIDIA’s H100 and B200-based platforms. It defines how compute trays, NVLink switch modules, and cooling infrastructure fit into a common rack frame, enabling scalable configurations like the NVL72. The architecture emphasizes modularity, allowing field crews to assemble, cable, and service components independently without disrupting adjacent modules.
Technical Details
Within an MGX rack, GPU trays connect to NVLink switch trays via a shared backplane or midplane using copper traces, eliminating external cables for intra-rack GPU-to-GPU traffic. Liquid cooling loops are routed through quick-disconnect couplings at each module, with coolant distribution manifolds integrated into the rack frame. The architecture supports both air-cooled and liquid-cooled variants, though Leviathan Systems primarily deploys the liquid-cooled versions for high-density clusters. Module insertion and removal use a standardized rail-and-latch mechanism with alignment guides—no torque screws are involved. Inter-rack scale-out connections (e.g., to other racks or the network) use fiber optic MPO connectors for high-bandwidth links.
How Leviathan Systems Works with MGX Modular Architecture
In the field, Leviathan crews assemble MGX racks by stacking compute and switch modules per the site-specific bill of materials, then connecting the internal liquid cooling loops and power backplanes. The modular design simplifies troubleshooting: a faulty GPU tray can be swapped without recabling the entire rack, which is common during commissioning or post-deployment maintenance.
Related service
GPU Rack Assembly & Rack-and-Stack →