华为全联接大会发布昇腾960超节点,全球首个采用NPO技术Huawei Connect releases Ascend 960 supernode, the world's first to adopt NPO technology.


AI free token/AI免费积分
华为全联接大会发布昇腾960超节点,全球首个采用NPO技术
0

今天上午,华为全联接大会2026在上海正式启幕。华为副董事长、轮值董事长汪涛发表《智启新将来,构建智能世界的硅基黑土地》主题演讲,现场发布全球首个搭载NPO工艺的昇腾960超节点,为十万亿级大模型训练与推理给出强大支撑。

汪涛表示,华为AI战略核心聚焦算力,深耕AI基础设施建设,致力于构建智能世界的硅基黑土地。面对大模型迈向十万亿规模的行业趋势,超节点已经成为超大规模AI基础设施建设的必然方向。华为围绕“超节点+集群”路线,依靠平台架构创新构筑算力竞争力,夯实本土算力底座,建设开源开放的算力生态,为全球AI发展给出全新挑选。

华为超节点以一套协议平等连接集群内全部组件,兼容跨服务器内存访问,使用近铜远光互连思路。本次大会推出全球首个量产NPO光引擎Hi-ONE,总容量7.2T,也是业内唯一内置光源的NPO机型,将多年积淀的光通信工艺融入超节点平台。

基于“灵衢+Hi-ONE”构建的昇腾960超节点正式亮相。作为全球首个应用NPO光引擎的超节点,机型使用正交架构搭配全液冷外观,依托灵衢达成内存统一编址,可扩展至4096卡,算力可达8EFLOPS FP8、16EFLOPS FP4。借助搭载5500个Hi-ONE,昇腾960超节点可替代4.8万颗800G光模块,平台功耗降低超550千瓦,无故障运行时间提高一倍,平台可用度高达99.8%,有力助推前沿大模型训练与推理落地。

与此同时,华为升级鲲鹏超节点,依托灵衢全光组网,最高兼容4096节点,构建256TB统一内存池,显著提高Agent沙箱密度、启动效率与向量检索能力。大会还发布OceanStor M900集群,基于灵衢UnifiedBus总线,达成一跳直连,给出L3.5层PB级KV缓存思路。多个昇腾960超节点可借助灵衢网络或RoCE互联,依托二层CLOS四平面组网,集群规模最高可达51.2万卡;叠加多轨道拓扑工艺,最大可支撑100万卡昇腾超节点集群。

算力生态建设同样是本次大会重点。汪涛提出,开源开放是华为算力核心战略,CANN已经达成全面开源开放,进入常态化社区运营阶段,昇腾生态已经跨越从“可用”迈向“易用”的关键拐点。

汪涛强调,华为将持续夯实AI基础设施,持续建设开源开放算力生态,兼容主流大模型在昇腾平台原生训练,拥抱百模千态,共赴智能世界加速到来的全新将来。

本文来源:https://www.163.com/dy/article/L730B28J05503FCU.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

0

This morning, the Huawei Connect 2026 conference officially kicked off in Shanghai. Huawei Vice Chairman and Rotating Chairman Wang Tao delivered a keynote speech titled "Enabling the New Future: Building Silicon-Based Fertile Ground for the Intelligent World," announcing the launch of the first-ever supernode powered by NPO technology, the Ascend 960, which provides robust support for training and inference of trillion-parameter-scale models.

Wang Tao stated that Huawei's AI strategy is centered on computational power, focusing on the construction of AI infrastructure to build a "silicon-based fertile ground" for the intelligent world. In response to the industry trend of models scaling up to trillions of parameters, supernodes have become the inevitable direction for large-scale AI infrastructure. Huawei is committed to the "supernode + cluster" roadmap, leveraging platform architectural innovation to build computational power competitiveness, consolidate the domestic computational base, and develop an open and collaborative computing ecosystem that offers a new choice for global AI development.

The Huawei supernode employs a unified protocol to connect all components within the cluster, supports cross-server memory access, and utilizes a "short copper, long fiber" interconnection strategy. The conference introduced the Hi-ONE, the world's first mass-produced NPO (Nonlinear Photonics) optical engine, with a total capacity of 7.2T. It is the only model in the industry that integrates the light source directly into the chip, bringing years of optical communication expertise to the supernode platform.

The Ascend 960 supernode, built on the "Lingqu + Hi-ONE" foundation, was officially unveiled. As the first supernode to utilize an NPO optical engine, it features an orthogonal architecture with a fully liquid-cooled exterior. Through the Lingqu system, it achieves unified memory addressing and can be scaled up to 4,096 cards, delivering a computational power of 8 EFLOPS for FP8 and 16 EFLOPS for FP4. By integrating 5,500 Hi-ONE modules, the Ascend 960 supernode replaces 48,000 800G optical modules, reduces platform power consumption by over 550 kW, doubles the mean time between failures, and achieves a platform availability of 99.8%, significantly accelerating the deployment of cutting-edge model training and inference.

Simultaneously, Huawei upgraded its Kunpeng supernode, which, thanks to the Lingqu all-optical networking, supports up to 4,096 nodes and creates a unified memory pool of 256 TB, significantly improving the density and startup efficiency of agent sandboxes as well as vector search capabilities. The conference also introduced the OceanStor M900 cluster, which utilizes the Lingqu UnifiedBus for single-hop direct connections and provides a PB-level KV cache based on a Layer 3.5 architecture. Multiple Ascend 960 supernodes can be interconnected via the Lingqu network or RoCE, forming a four-plane CLOS network capable of supporting up to 512,000 cards. When combined with multi-rail topology, the system can handle a maximum of a million cards.

The development of the computing ecosystem is another key highlight of this conference. Wang Tao noted that open-source collaboration is a core strategy for Huawei's computing power, with CANN already fully open-sourced and entering a phase of regular community operation. The Ascend ecosystem has successfully transitioned from being "usable" to "easy to use."

Wang Tao emphasized that Huawei will continue to strengthen its AI infrastructure and develop an open and collaborative computing ecosystem. The company aims to support the native training of mainstream models on the Ascend platform, embracing the diverse landscape of AI models to usher in a new era where the intelligent world arrives faster.

Huawei Connect releases Ascend 960 supernode, the world's first to adopt NPO technology.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry