H3C releases UniPoD S81000 X1 super node
新华三发布 UniPoD S81000 X1 超节点，单域支持 40 张国产 GPU
This product addresses critical bottlenecks in large AI model training with its all-to-all communication architecture, reducing data waiting times significantly. For enterprises in ASEAN deploying trillion-parameter MoE models or running 24/7 AI agents, the liquid-cooled, cable-free design offers 10x physical reliability and linear scalability without proportional cost increases.
H3C launched the UniPoD S81000 X1 super node on 29 Sep 2023.
The rack-level fully interconnected AI computing product supports 40 domestic GPUs in a single Scale-Up domain.
It achieves 448GB/s P2P bandwidth between any two cards through a CLOS topology.The product aggregates 5760GB of unified addressable memory and delivers 28P FLOPS of computing power at FP8 precision.
新华三发布 UniPoD S81000 X1 超节点，单域支持 40 张国产 GPU
IT之家 9 月 29 日消息，新紫光集团旗下新华三昨日发布了机架级全互联智算产品 UniPoD S81000 X1 超节点。该产品单 Scale-Up 域最大支持 40 张国产加速卡，通过 4 个 Scale-Up 专用交换 Tray 实现一级 CLOS 全互联拓扑，任意两卡间 P2P 带宽达 448GB/s。 S81000 采用正交连接设计，计算节点与交换节点通过高速连接器直连，无须线缆与背板。产品采用纯电互联，无光模块，厂商称物理可靠性提升 10 倍。 显存与算力方面，40 卡聚合为 5760GB 可统一编址的显存池，支持 FP8 精度（IT之家注：FP8 为 8 位浮点格式，常用于 AI 训练与推理加速），总算力可达 28P FLOPS。 产品兼容 19 英寸标准机柜，支持机柜解耦交付，单柜最大可部署 2 套 40 卡超节点。散热采用液冷为主、风液兼容的协同架构。 传统 8 卡服务器依靠以太网堆叠组网，卡间互联停留在微秒级时延、百 Gb/s 带宽。万亿参数 MoE 大模型的 All-to-All 通信需求下，大量 GPU 时间耗费于数据等待。此外，智能体 7×24 小时运行，多 Agent 串联复合失败率显著。传统集群组件繁杂、故障定位缓慢，运维成本随规模线性增长。 新华三表示，随着 AI 智能体加速渗透企业工作流，智能体任务对底层计算资源形成三重压力。单次任务 Token 消耗达十万至百万级，较传统对话场景增长数十倍。
This report is categorized under AITOP's "Products" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include H3C, UniPoD S81000 X1, GPU clusters, AI infrastructure.
Source attribution: curated from IT之家.
