According to a research report from Orient Securities, super nodes serve as the "spearhead" for the advancement of domestic computing power, compensating for the performance limitations of individual cards through the super node approach. The year 2026 is identified as the inaugural year for the mass production of domestic super nodes.
The integration of a Scale-up domain within super node cabinets generates demand for switch chips and switching nodes. High-speed interconnections are required between switching nodes and compute nodes, fostering needs for copper cables, PCB backplanes, and optical interconnects. The significant increase in power consumption of super node cabinets makes liquid cooling an essential requirement, signaling a gradual transition to a fully liquid-cooled era. The complexity of super node cabinets, involving multiple components, raises industry entry barriers and increases value-retention segments, leading to a reassessment of server ODM manufacturers. The continuous growth in super node power consumption is restructuring power supply architectures, creating demand for PSU, HVDC, and SST.
Key viewpoints from Orient Securities are as follows: AI development necessitates a comprehensive shift of AI infrastructure towards the super node era. While some investors view super nodes merely as an escalation from eight-card to multi-card computing power stacks, the firm contends that super nodes are an essential choice for the evolution of computing infrastructure on both the training and inference sides, forming the foundation for future million-card clusters.
(1) Training Side: As model parameters increase and the MoE architecture evolves, TP and EP create demands for high-bandwidth, frequent interconnections. Super node servers, utilizing internal high-speed bus interconnects, can effectively support parallel computing tasks and reduce the training cycle for large models. (2) Inference Side: In the era of "raising lobsters," inference tasks place high demands on KV Cache. Super nodes effectively address the "memory wall" issue and offer higher tokens generated per watt per second, providing greater cost-effectiveness for inference. (3) Million-Card Clusters: Scale-up initiatives are leading the way, with super nodes acting as the foundational base for future large-scale clusters.
Scale-up protocols are moving towards open-source and open standards. Super nodes facilitate a "corner overtaking" strategy for domestic chips. Super nodes can be categorized into full-rack super nodes, split-rack super nodes, and Matrix super nodes (cascaded super nodes). They primarily consist of compute nodes, switch nodes, TOR switches, power supply units, power busbars, cable trays/orthogonal backplanes, and liquid cooling components. Manufacturers, led by Broadcom and Alibaba, are progressively initiating open standard protocols, breaking the relatively closed competitive landscape of Scale-up protocols and accelerating the development of the super node ecosystem.
Although domestic chips slightly lag in manufacturing processes—for instance, a single Ascend 910C chip's BF16 performance is only one-third that of a GB200 module, as seen in the Huawei 384 versus GB200 NVL72 comparison—a 384 cluster utilizing the super node approach achieves an overall BF16 performance 1.7 times that of the NVL72.
Super node servers introduce five major trends: (1) The addition of a Scale-up domain in super nodes, with its continuous expansion, drives an increasing ratio of switch chips. (2) The need for high-speed interconnection between switch nodes and compute nodes, alongside the growing number of computing cards, gradually spurs demand for PCB backplanes and orthogonal connectors. (3) The trend towards higher density in super node cabinets and continuously rising power consumption fuels the need for 100% full liquid cooling. (4) The transition from 8-card servers to super nodes involves multiple aspects such as networking, power supply, and cooling, raising industry entry barriers, increasing value-retention segments, and leading to a reassessment of server ODM manufacturers. (5) Super nodes are restructuring power supply architectures, generating demand for PSU, HVDC, and SST.
Risk warnings include AI development falling short of expectations, intensifying industry competition, and geopolitical risks.
Comments