One Architecture. Every Scale.
MGX — The Modular GPU X reference design for AI training and inference. Base-8 to Base-24 connectivity. 800G to 3.2T lane speeds. Rail-optimized, front-access, and built to grow.
Informational Reference
NVIDIA / OCP Ecosystem
Modular by Design.
Optical by Necessity.
MGX (Modular GPU X) is a reference architecture for deploying GPU-intensive compute in standard rack form factors. It defines the sled, the rack layout, and the network topology — but not the cabling. A typical MGX rack houses 9 AI nodes, each with 8 GPUs, plus leaf switches positioned between the nodes. The GPUs communicate chip-to-chip via high-bandwidth links; the nodes communicate rack-to-rack via optical spine networks. As lane speeds rise from 800G to 1.6T and toward 3.2T, the physical layer shifts from copper to co-packaged optics to external fiber — making the optical connectivity strategy a defining factor in cluster performance.
Core Architectural Principles
Match the Transceiver to the Topology
MGX supports Base-8 (SR8/DR8), Base-16, and Base-24 MPO/MTP® connectivity. Base-8 dominates 800G deployments. Base-16 and Base-24 emerge at 1.6T and 3.2T to reduce connector count and simplify cable management without sacrificing lane density.
One Failure. Not One Halt.
In rail-optimized architectures, the 8 lanes of an 800G or 1.6T link are distributed across independent switch rails. A single transceiver failure drops bandwidth by 12.5% or 25% — not 100%. The cluster continues training. Maintenance waits for your schedule, not your job queue.
Keep the Front Face Free
MGX sleds reserve the front face for ELS pluggable modules, loopback testing, and thermal management. All high-fiber-count trunking migrates to the rear via blind-mate or pass-through fiber busbars. This preserves front-face airflow and serviceability while managing massive spine fiber counts at the back.
Move the Optics Closer
Co-Packaged Optics (CPO) and Near-Package Optics (NPO) place the optical engine on or near the ASIC, reducing power and latency versus traditional pluggable transceivers. MGX accommodates both: front-face ELS pluggables for flexibility, and rear-mounted passive fiber cassettes for CPO backplane integration.
9 Nodes. 72 GPUs. Spine Outside.
Rack Layout
- 1 MGX rack houses 9 AI nodes
- Each node = 8 GPUs (4 top, 5 bottom)
- Leaf switches sit between the AI nodes
- 8× GPU C2C (chip-to-chip)
- 2× Front End
- 1× In-Band Management
- 144 ports per 2RU
- 72 ports to GPUs (downstream)
- 72 ports to Spine network (upstream, outside the rack)
Generic rack elevation · rear fiber paths · no vendor logos
Why the Optical Layer Matters
Nanoseconds Cost Millions
Distributed AI training requires nanosecond-level synchronization across thousands of GPUs. The propagation delay of the fiber itself — not just the transceivers — becomes a bottleneck. Lower-latency fiber technologies and shorter, pre-validated optical paths directly improve model convergence time.
1,000 Fibers per Rack
A fully populated MGX rack can require over 1,000 fiber lanes between GPUs, leaf switches, and spine uplinks. Managing this with traditional patch cords creates cable nests that block airflow and complicate MACs. Rail-optimized shuffle architectures and rear-mounted fiber busbars replace manual patching with factory-validated paths.
Train Through Failure
In standard link-to-link fabrics, one failed transceiver halts the entire training job. Rail-optimized shuffle architectures distribute lanes across independent rails so that single failures result in partial bandwidth reduction, not catastrophic loss. The physical layer protects the workload.
Architecture Governance
Connecting the MGX Layer?
ADTEK supplies the rail-optimized shuffle cassettes, rear-mounted fiber busbars, and blind-mate optical backplanes that integrate with MGX and ORv3 rack architectures. Discuss your GPU cluster connectivity strategy with our engineering team.
