中科曙光高级副总裁李斌先生在演讲中也指出，随着大模型技术的持续地迭代发展，对计算装备提 出了前所未有的新要求和新挑战。在他看来，这些挑战主要集中在内存的容量和带宽、线间互联的 带宽和延迟、节点间的互联带宽和延迟、系统的能耗和能源效率、系统的稳定性以及应用生态的兼 容性这几个方面。 面对这些挑战，李斌认为，我们需要从两个方面突破：一是要从单点突破到系统工程；二是要合作 开放。"在此基础上，我们把系统的冷却供电和管理这些方面也进一步耦合来提效，同时在这个过 程中，我们开放共享，去促进跨层的软硬件的协同和产业链的合作共赢。"李斌接着说。 如果您希望可以时常见面，欢迎标星收藏哦~ 日前，首届光合组织人工智能创新大会（HAIC2025）在江苏昆山隆重开幕。全产业链超2500家企 业代表、专家学者及行业领袖，通过200余场高密度分享与5000余平方米实景展区，全景呈现中国 AI计算开放架构的创新实践与生态繁荣。 从这次的大会分享中，我们看到了一个属于中国AI的新时代，正在缓缓走来。 摩尔定律放缓下的取舍 熟悉人工智能的读者对Scaling Law、Scale Up和Scale Out等概念应该非常了解。在笔者看来，这 些 ...

同样在今年，华为昇腾384超节点真机在2025世界人工智能大会期间首秀。从华为昇腾384超节点到 scaleX万卡超集群接连真机亮相，国产算力"大基建"正从图纸走进现实。 超节点支撑万卡级集群 随着主流大模型参数从千亿级跃升至万亿级规模，算力需求持续爆发式增长，EFLOPS算力级别、万卡 级别高性能集群成为大模型标配。 超节点（SuperPod）因密度和性能优势，正成为新建万卡集群的首选架构之一。 其应算力需求而生。中科曙光高级副总裁李斌在接受21世纪经济报道等媒体采访时指出，在算力要求不 高的时代，IT组件的解耦设计提供了无与伦比的灵活性，允许通过标准网络轻松扩展。然而，当前算法 对算力的苛刻要求，迫使业界必须牺牲部分通用性，以换取性能的极致优化。 在此之前，国内的万卡集群项目多以分散服务器、技术蓝图或在建工程的形式存在。如三大运营商已投 产运营的万卡集群，多为分散部署在庞大机房内的标准服务器，虽然已实现商业运营，但其物理形态并 非一体化单元。 scaleX万卡超集群由16个曙光scaleX640超节点互连而成。作为世界首个单机柜级640卡超节点， scaleX640采用超高密度刀片、浸没相变液冷等技术，将 ...

会议上，中科曙光就市场关注的超节点640产品订单、适配H200及与英伟达合作等问题作出回应。中科曙光方面 表示，曙光scaleX640依托先进的硬件架构、冷却、供电等技术，集约化程度全球领先，超节点综合性能国内领 先。随着大模型参数规模进一步增大和高通量推理集群化部署，该产品优势将更加突出，具备广阔市场前景。此 外，曙光scaleX640超节点采用AI计算开放架构，支持市场上主流的国产和国际AI加速卡，用户可按需选择加速 芯片，实现"软硬协同、生态兼容"的AI计算解决方案。 此外，伟创电气、福莱新材、长安汽车均具有"机器人"概念，近一个月机构来访接待量分别为141家、135家、 125家。 智通财经12月14日讯（编辑 若宇）据Choice数据显示，截至发稿，近一个月（11.23-12.14）包括杰瑞股份、中科 曙光、海光信息、伟创电气、福莱新材、长安汽车、乐鑫科技、恺英网络、世纪华通、纳科诺尔、三峡旅游在内 的11家上市公司机构来访接待量超90家（含90家），具体情况见下图： | | | 机构来访接待量(11.23-12.14) | | | | --- | --- | --- | --- | --- | | ...

在本次终止重大资产重组投资者说明会上，海光信息高管指出，交易双方的二级市场股价相比预案时发 生了较大变化。特别是自8月中旬以来，受国内国际环境变化、A股市场整体走势、AI产业热度变化及 市场预期等复杂因素的影响，吸并双方股价开始整体上涨并出现较大波动。基于审慎性考虑，决定终止 本次交易事项。 随着海光信息吸收合并中科曙光事项按下"终止键"，两家国产算力巨头将在各自赛道上加速起航。12月 10日，海光信息、中科曙光分别召开终止重大资产重组投资者说明会，双方高管还同时出席了当天线上 交流会。两家公司高管都指出，股价上涨对重组终止产生了影响，否认前期信息披露不充分，强调后续 双方将在保持上市公司独立性的基础上加强战略协同。 股价变化较大 后，12月10日海光信息微跌，报收218.5元/股，中科曙光股价跌停，收于90.12元/股。 本次说明会上，还有投资者提问中科曙光高管是否有减持海光信息的计划。公司高管表示，结合目前公 司发展实际，持股并深化产业协同更符合双方长期利益；后续是否有减持计划结合自身业务战略、资金 状况、产业协同性及行业发展趋势综合判断。 否认风险披露不充分 在重组说明会上，不少投资者提出"11月29日 ...

北京商报讯（记者 陶凤 王天逸）12月10日，中科曙光披露投资者关系活动记录表，就市场关注的超节 点640产品订单、适配H200及与英伟达合作等问题作出回应。 中科曙光方面表示，曙光scaleX640依托先进的硬件架构、冷却、供电等技术，集约化程度全球领先， 超节点综合性能国内领先。随着大模型参数规模进一步增大和高通量推理集群化部署，该产品优势将更 加突出，具备广阔市场前景。此外，曙光scaleX640超节点采用AI计算开放架构，支持市场上主流的国 产和国际AI加速卡，用户可按需选择加速芯片，实现"软硬协同、生态兼容"的AI计算解决方案。 ...

Core Insights - The leading domestic large-scale intelligent computing supercluster system is set to debut at the HAIC2025 from December 17 to 19, 2025, in Kunshan, which will provide comprehensive support for breaking through the domestic AI computing power bottleneck and further improve the AI computing power industry chain layout [1][2] Group 1: Technology and Design - The supercluster system is designed with an open collaboration core concept and will feature domestically developed high-speed interconnect network technology, which is expected to break industry records for computing power density and achieve both "scale expansion" and "performance enhancement" [1] - The system is based on the Shuguang scaleX640 super node technology, potentially reaching a computing power scale of tens of thousands of cards, placing it at the forefront of domestic intelligent computing devices [2] Group 2: Performance and Compatibility - The scaleX640 super node is the world's first single cabinet-level 640-card super node, with a "one-to-two" high-density architecture that can form a thousand-card computing unit, doubling performance and increasing density by 20 times compared to similar products [2] - In scenarios involving MoE trillion-parameter large models, performance is expected to improve by 30% to 40%, and the system is compatible with multiple brands of domestic accelerator cards, showcasing significant adaptability and cost-performance advantages [2] Group 3: Industry Impact - The potential release of the tens of thousands of card supercluster represents a milestone for domestic intelligent computing, transitioning from "single-point breakthroughs" to "clustered implementation" [2]

Core Viewpoint - The computer industry index has underperformed compared to major stock indices, indicating a challenging market environment for the sector [1][2]. Market Review - During the week of November 3 to November 7, the Shanghai Composite Index rose by 1.08%, the ChiNext Index increased by 0.65%, and the CSI 300 Index gained 0.82%. In contrast, the computer (Shenwan) index fell by 2.54%, lagging behind the Shanghai Composite by 3.62 percentage points, the ChiNext by 3.19 percentage points, and the CSI 300 by 3.36 percentage points, ranking 30th among all industries [1][2]. Weekly Insights - NVIDIA is leading the trend of supernodes, a technology architecture for building large-scale computing clusters, which integrates thousands of GPUs into a single logical unit. The latest NVLink technology has reached its fifth generation, with each GPU having 18 NVLink connections, achieving a total bandwidth of 1800GB/s, significantly surpassing PCIe Gen6 [3]. - NVIDIA's upcoming NVL72, set to be released in March 2024, will integrate 36 Grace CPUs and 72 Blackwell GPUs into a liquid-cooled cabinet, delivering a total of 720 PFLOPs for AI training and 1440 PFLOPs for inference [3]. Domestic Major Players Accelerating Supernode Layout - **Inspur**: On November 6, during the World Internet Conference, Inspur launched the world's first single-cabinet 640-card supernode, achieving a 20-fold increase in computing density [4]. - **Huawei**: In April, Huawei introduced the CloudMatrix384 supernode, capable of creating a super-large cluster with over 160,000 cards. As of September, over 300 units have been sold, primarily to government and enterprise clients [4]. - **Alibaba**: At the 2025 Cloud Computing Conference, Alibaba Cloud unveiled the Panjiu 128 supernode AI server, which enhances inference performance by 50% compared to traditional architectures [5]. - **Baidu**: Announced the launch of the Kunlun supernode at the 2025 Baidu Cloud Intelligence Conference, making supercomputing capabilities available [5]. - **ZTE**: Developed a supernode server with 64 GPUs, featuring an innovative design that reduces latency to the nanosecond level [5]. - **Inspur Information**: Released the "Yuan Nao SD200" supernode AI server aimed at trillion-parameter models [5]. Investment Recommendations - Focus on companies involved in computing power such as Cambricon, Haiguang Information, Inspur, and others [6]. - Consider AIDC-related firms like Kehua Data and Yunse Intelligent [6]. - Explore AI application companies including Kingsoft Office, iFlytek, and others [6].

Core Insights - The company has officially launched the world's first single cabinet-level 640-card super node, scaleX640, which was showcased at the Wuzhen Internet of Things Expo [1] - The scaleX640 super node features a "one-to-two" high-density architecture, achieving ultra-high-speed bus interconnection within a single cabinet, and can form a thousand-card computing unit through dual-node combinations [1] - Compared to similar products in the industry, the scaleX640 has doubled its overall computing performance and increased cabinet computing density by 20 times [1] - In applications such as MoE trillion-parameter model training and inference, the system-level performance improves by 30%-40% compared to traditional solutions [1] - The company maintains a positive outlook on its technological leadership and product competitiveness in the intelligent computing infrastructure sector, with net profit forecasts for 2025-2027 at 2.652 billion, 2.984 billion, and 3.349 billion yuan, representing year-on-year growth of 38.8%, 12.5%, and 12.2% respectively [1] - Valuation-wise, the company is compared to server manufacturers, with a target price of approximately 126 yuan based on a 62x PE valuation for 2026, maintaining a "strong buy" rating [1]

Core Viewpoint - The year 2025 is anticipated to be a breakthrough year for domestic supernodes, with major companies like Inspur, ZTE, Huawei, Alibaba, and Sugon making significant advancements in computing cluster construction, enhancing computing power integration, density, and ecosystem compatibility [2]. Group 1: Product Developments - Huawei's Ascend 384 has set a new industry standard as the largest high-speed bus interconnected supernode, featuring 32 cards per cabinet across 12 cabinets, showcasing Huawei's comprehensive capabilities in communication and computing [2]. - Alibaba's Panjiu AL128 supernode has achieved a record of supporting 128 accelerator cards in a single cabinet, with a computing power integration level four times that of the Ascend 384, demonstrating rapid advancements in software and hardware optimization [2]. - The Sugon scaleX640 supernode is the world's first single-cabinet 640-card supernode, achieving 20 times the computing power integration of the Ascend 384, designed on an open AI computing architecture to ensure compatibility with mainstream intelligent computing ecosystems [2]. Group 2: Performance Comparison - Domestic supernodes have undergone three significant leaps, overcoming barriers in performance and ecosystem, with the scaleX640 showing core advantages over NVL72 in comprehensive performance metrics [3]. - The scaleX640 has implemented advanced immersion phase change liquid cooling technology, achieving a minimum PUE of 1.04 and providing 1.72MW of cooling capacity for high-caliber computing units, validated through over 30 days of reliability testing [3]. - Despite a gap in single-card computing power compared to NV, the engineering characteristics of computing clusters present systemic opportunities for domestic manufacturers to catch up, with ongoing innovations in integration, compatibility, and reliability [3].