Analysis of Huawei Ascend 384 Super Node Technology and Industry Chain Overview
I. Breakthroughs in Ascend 384 Super Node Technology and Its Significance for the Industry
At the World Artificial Intelligence Conference held on July 26, 2025, Huawei publicly showcased its latest research and development achievement: the Ascend 384 super node computing system. This system consists of 12 standard computing cabinets and four dedicated bus cabinets, innovatively achieving a high-speed interconnection architecture with 384 NPU computing cards. From a technical parameter perspective, this system has made significant breakthroughs in computing density, interconnect bandwidth, and energy efficiency ratio, earning it the honorary title of "Treasure of the Museum" from the conference organizing committee.
Compared to international counterparts like NVIDIA's NVL72 solution, the Ascend 384 super node demonstrates remarkable advantages across multiple performance metrics. Firstly, regarding computational density, through innovative bus architecture design, single cabinet computational output is increased by approximately 35%; secondly, in terms of interconnect bandwidth, using Huawei's self-developed ultra-fast interconnect protocol reduces communication latency between nodes to nanosecond levels; finally, thanks to deep optimization of liquid cooling systems for energy efficiency performance control within an industry-leading PUE value below 1.15.
This technological breakthrough holds strategic significance for China's artificial intelligence infrastructure construction. The launch of the Ascend 384 super node signifies that China possesses a complete technology stack that is independently controllable in high-end AI computing clusters which provides a solid computational foundation for intelligent transformation across various industries. Particularly in computation-intensive scenarios such as large model training and scientific calculations as well as smart cities deployment at scale will effectively alleviate current industry pain points related to insufficient AI computational supply.
II. Computing Power Scheduling and System Integration Ecosystem
As a key supporting aspect of super node systems，the field of power scheduling and system integration has formed a complete industrial ecosystem。Hengwei Technology，as a leading enterprise in this field，has successfully applied its developed intelligent operation maintenance system into China Mobile’s AI power scheduling platform project with winning bid amount reaching CNY300 million。This system can achieve dynamic allocation和load balancingof cross-regional power resources through intelligent resource distribution algorithms。
Zhizhen Technology focuses on optimizing communication protocols within ascend AI compute pools，其研发的节点间通信加速技术显著降低了分布式训练过程中的数据同步延迟。在系统集成服务方面，神州数码作为华为昇腾产品线的核心合作伙伴已承接武汉超节点建设项目累计采购规模达8．6亿元，该项目涉及从硬件部署到软件调优的全生命周期服务。
值得注意的是软通动力凭借其在大模型算法调优领域的技术积累，为包括科大讯飞、商汤科技在内多家AI头部企业提供算力优化服务。而东方国信参与上海算力调度平台建设项目则开创性地实现长三角区域算力资源跨域协同调度，为区域一体化数字经济发展提供新型基础设施支撑。
III.Server Manufacturing Supply Chain Analysis
In server manufacturing环节中华鲲振宇（高新发展控股）作为昇腾910B整机主力供应商2025年第一季度营收同比增长172%反映出市场对国产AI服务器强劲需求。其产品线覆盖从训练型到推理型全场景服务器解决方案，在能效比和可靠性方面具有显著优势。 yunhuo Communication作为华为Atlas系列服务器定制化供应商专门为超节点系统开发高密度机柜解决方案通过创新模块化设计单个42U标准机柜可集成多达32个计算节点空间利用率提升40%以上拓维信息深度参与区域算力中心建设其长沙部署超节点集群已投入商业运营支持当地智能网联汽车等新兴产业的发展。 yuqianglong将军用散热技术成功转化民用液冷解决方案成为贵阳超节点项目独家机柜供应商开发相变液冷系统可实现单机柜50kW以上散热能力同时噪音控制65分贝以下工业富联全球AI服务器代工龙头为超节点系统提供关键结构件支持精密制造工艺确保大批量生产时质量一致性。 yunyuan_xinxin光通信与高速互联技术光通信模块作为超节点系统神经网络中际旭创已实现1．6T光模块批量供货创新光电协同设计功耗降低30%。 xinyisheng同步推进800G与1．6T产品研发迭代独创封装工艺使得良率稳定99．5%以上光迅科技CPO（共封装光学）领域取得突破通过芯片级光电集成传输密度提升5倍。在高速连接器领域华丰科技研发56Gbps背板连接器打破国外垄断独特信号完整性设计串扰降低-50dB以下瑞可达开发液冷快换接口系统支持热插拔操作单个接口更换时间不超过3分钟，大幅提升运维效率意华股份PCIe5．0连接器方案通过华为认证阻抗控制精度达到±3%的行业顶尖水平。Advanced Cooling System Innovations highland Co.'s immersion liquid cooling solution occupies60% domestic market share，其单集群解决方案价值1200万元该技术通过全浸没设计将芯片结温控制70℃以下，同时PUE值1．08极致能效网宿科技开发冷板式液冷采用分区温控策略根据负载动态调整流量节能效果25%以上。在散热核心部件方面强瑞技术密封组件漏液率低于0．001%的行业纪录多层复合结构承受10万次循环测试科创新源石墨烯导热材料导热系数1500W/mK传统材料五倍飞荣达相变导热材料成功应用NPU芯片散热接触阻抗低至0·05℃·cm²/W。### V.Progress on Chip Materials & Storage Technologies In semiconductor materials domain博威合金研发高导铜合金替代美国Materion产品，其 thermal conductivity reaches400W/mK，与 chip substrate perfectly matched。这种材料已经用于NPU heat dissipation substrates production with yield rate stabilizing98%雅克科技前驱体突破国外限制Low-α射线 material可以将soft error rate reduced one order magnitude storage配套江波龙企业级SSD自研主控芯片supports PCIe5 .0 interface 和NVMe protocol random read/write performance reaches3000KIOPS兆易创新DDR5 memory chips经过Ascend memory pool certification其innovative Bank Group架构有效带宽提高85%包装基板兴森科技IC载板应用于Ascend910B packaging微孔加工精确达到±15μm industry leading level ### VI.Software Development & Ecosystem Construction润和软件developed MindSpore framework optimization tool can improve large model training efficiency20%. Its innovative automatic parallel strategy supports distributed training at thousand-card scale gradient synchronization efficiency reaches95%.常山北明针对政务场景developed specialized large models have been successfully deployed on supernode power platforms supporting over20 types including document generation policy interpretation etc 东方通R&D distributed data scheduling middleware achieves cross-node data access delay lower than five milliseconds while its smart prefetch algorithm improves data hit rates up90%.中国软件完成麒麟OS对Ascend platform compatibility adaptation covering full-stack optimizations from driver layer application layer shortening response time40%. These software ecosystems provide critical support for commercial applications.