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机型 UniPoD S81000 X1 超节点。该机型单 Scale-Up 域最大兼容 40 张国产加速卡,借助 4 个 Scale-Up 专用交换 Tray 达成一级 CLOS 全互联拓扑,任意两卡间 P2P 带宽达 448GB/s。
S81000 使用正交连接外观,计算节点与交换节点借助高速连接器直连,无须线缆与背板。机型使用纯电互联,无光模块,厂商称物理可靠性提高 10 倍。
显存与算力维度,40 卡聚合为 5760GB 可统一编址的显存池,兼容 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。
机型兼容 19 英寸标准机柜,兼容机柜解耦交付,单柜最大可部署 2 套 40 卡超节点。散热使用液冷为主、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 通信需要下,大量 GPU 时间耗费于数据等待。另外,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,较传统对话情形增长数十倍。
本文来源:https://www.ithome.com/1/008/396.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Thanks to IT之家 netizen "不一样的感受" for the tip!
IT Home News, September 29: H3C, under New Unigroup, yesterday released the rack-level fully interconnected AI computing model UniPoD S81000 X1 super node. This model supports up to 40 domestically produced accelerator cards in a single Scale-Up domain, and with 4 dedicated Scale-Up switch trays achieves a one-level CLOS fully interconnected topology, with P2P bandwidth of 448GB/s between any two cards.
The S81000 uses an orthogonal connection design; compute nodes and switch nodes are directly connected via high-speed connectors without cables or backplanes. The model uses purely electrical interconnection with no optical modules; the manufacturer claims physical reliability is improved 10-fold.
In terms of memory and compute, the 40 cards aggregate into a 5760GB uniformly addressable memory pool, supporting FP8 precision (IT之家 note: FP8 is an 8-bit floating-point format commonly used for AI training and inference acceleration), with total compute reaching 28P FLOPS.
The model is compatible with 19-inch standard racks and supports rack-decoupled delivery; up to 2 sets of 40-card super nodes can be deployed per rack. The cooling uses a collaborative architecture with liquid cooling as the primary method, compatible with both liquid and air cooling.

Traditional 8-card servers rely on Ethernet stacking for networking, with inter-card interconnect stuck at microsecond-level latency and hundreds of Gb/s bandwidth. Under the All-to-All communication demands of trillion-parameter MoE large models, large amounts of GPU time are consumed waiting for data. In addition, agents run 7×24 hours, and the compound failure rate of multiple Agents connected in series is significant. Traditional clusters have complex components and slow fault localization, with O&M costs growing linearly with scale.
新华三 stated that as AI agents accelerate their penetration into enterprise workflows, agent tasks place triple pressure on underlying computing resources. A single task consumes 100,000 to 1,000,000 Tokens, tens of times more than traditional conversation scenarios.