Ruijie RG-ETH 128 Supernode: Dual Architecture Scales AI Clusters to 1024 GPUs

The Ruijie RG-ETH 128 supernode (Supernode 1.0) adopts a dual-architecture design combining compute nodes and switch nodes, integrated at the cabinet level to achieve high-speed interconnect for 64 or 128 GPUs.

Compute Node Architecture

Each compute node fits in 1U and integrates 1 CPU with 4 GPU OAM modules interconnected via a PCIe Switch. Two deployment modes are offered:

Non-decoupled (default): Each compute tray contains 1 CPU + 4 GPUs, yielding 64 GPUs per full cabinet.

Decoupled (Host Tray): CPUs are separated into dedicated Host Trays, freeing space for 8 GPU OAM modules per compute tray, doubling density to 128 GPUs per cabinet.

Cooling uses a hybrid approach: cold-plate liquid cooling for CPUs, GPUs, and power conversion modules; air cooling for lower-power components.

Switch Node Design

The switch node is built around a 51.2 Tb/s high-performance switch ASIC (TH Ultra) with 250 ns ultra-low latency, supporting the SUE protocol and reliability features such as LLR and CBFC. Its 51.2 Tb/s bandwidth is split equally:

25.6 Tb/s via Cable Tray to all GPUs for intra-cabinet scale-up (TYPE1 closed cabinet interconnect).

25.6 Tb/s via Fly Cable to 32 front-panel OSFP-800G ports for scale-up expansion, enabling multi-cabinet clustering and flexible HBD scaling from 64 to 1024 GPUs (TYPE2 intra-cabinet + cross-cabinet).

Two Switch Node Variants

TYPE1 (closed): No front-panel ports; all 51.2 Tb/s connects to rear connectors, saving 50% switch ASIC cost.

TYPE2 (expandable): 32×800G front-panel expansion ports + 25.6 Tb/s rear Cable Tray for cross-cabinet growth.

Scale-Up Evolution Path

The article outlines a clear two-step roadmap for larger scale-up domains:

Copper evolution: SerDes upgrades from 112G to 224G, reducing cable density and improving yield.

Optical interconnect: When scale-up domains reach 512–1024 GPUs, optical interconnect becomes mandatory, progressing through LPO → NPO → CPO module technologies.

Cabinet Interconnect Implementation

Achieving 128 GPUs per cabinet relies on two complementary schemes:

Host Tray decoupling physically separates CPUs from GPU trays, allowing 8 GPUs per 1U tray and doubling single-cabinet density.

Supernode cabinet clustering uses the 32×800G expansion ports on TYPE2 switch nodes to link multiple cabinets into a larger high-bandwidth domain (HBD), officially supporting 64 to 1024 GPU scaling.

Accompanying architecture diagrams illustrate the compute/switch node layout, bandwidth allocation, TYPE1 vs TYPE2 switch configurations, and the Host Tray decoupling mechanism.