China Mobile's Open Disaggregated Supernode (ODS) Architecture for AI Compute

China Mobile Cloud proposes the Open Disaggregated Supernode (ODS) architecture to overcome the closed, tightly coupled design of traditional AI supernodes. The article first outlines five core challenges of legacy supernode architectures:

Hardware tight coupling : GPU, switching, power, and cooling are integrated in a closed chassis, preventing independent component replacement or upgrade.

Proprietary network protocols : Vendor-specific interconnect protocols make multi-vendor GPU and switch interoperability difficult, locking users into a single supplier.

Power and cooling pressure : Single-rack power density surges beyond the capability of traditional enclosed cooling solutions.

Limited system scalability : Vertical scale-up hits a ceiling; horizontal scale-out incurs high cost.

Poor interconnect maintainability : Complex internal cabling complicates fault isolation and repair.

The ODS architecture targets five design goals: full hardware disaggregation, fully open network protocols, elastic system scaling, flexible cooling options, and optimized interconnect efficiency.

Primary Node Forms

Compute Node (4U)

Supports both air-cooled (4 GPUs) and liquid-cooled (8 GPUs) configurations. Features include FP8 compute, high HBM bandwidth, copper interconnects (DAC/AEC), and optical modules (FRO/LPO/LRO). Network protocols support standard Ethernet plus the IETF draft FARE-in-SUN adaptive routing protocol.

Switch Node (4U Ethernet Switch)

Supports 400G QSFP112 / 800G OSFP ports. One device serves multiple roles: scale-up, scale-out, and front-end networks. Implements RoCE and FARE full adaptive routing Ethernet protocol.

Two ODS Deployment Forms

1. Basic ODS

Air-cooled max 64 GPUs; liquid-cooled max 128 GPUs. Uses high-speed copper DAC/AEC for interconnect. Achieves ~50% power reduction, ~50% cost reduction, and two orders of magnitude reliability improvement. Completed multi-vendor 800G AEC/DAC signal integrity testing.

2. Cascaded ODS (with OCS Optical Circuit Switch)

Multiple basic ODS pods interconnected via a 256×256 OCS to build larger compute pools. Delivers ~90% power reduction, ~60% cost reduction, and further reliability gains. Completed OCS insertion loss and switching error validation. Provides a smooth evolution path toward next-generation co-packaged optics (CPO) while avoiding CPO's current maintainability, cooling, and yield challenges.

Industry Development Trends

The article surveys three technical routes across the industry:

Open protocol route : Microsoft Maia200, AMD Helios (UALoE), Meta MTIA series.

Hardware disaggregation route : NVIDIA MGX ETL, Huawei CM384, Tencent Eth-X Ultra, AMD MI500 UAL256, Alibaba UPN512.

Optical circuit switching / hybrid optical-electrical switching : Google TPUv7 Ironwood, NVIDIA NVL576/NVL1152, Microsoft Maia200 (6,144-card scale).

On the ecosystem front, China Mobile Cloud leads the Open Data Center Committee (CODCC) to drive open disaggregated supernode standardization, with participation from Intel, Broadcom, Inspur, Suiyuan, Biren, Hygon, Sugon, FiberHome, and dozens of other companies.