Core Viewpoint - Tianshu Zhixin (09903) has announced a roadmap for its fourth-generation chip architecture, aiming to surpass NVIDIA's Rubin architecture by 2027, which has positively impacted its stock price, increasing by 4.26% to HKD 178.5 with a trading volume of HKD 139 million [1] Group 1: Chip Architecture Roadmap - The company plans to exceed NVIDIA's Hopper architecture with its Tianshu Tianshu architecture by 2025 [1] - The Tianshu Tianxuan architecture is set to match NVIDIA's Blackwell architecture in 2026 [1] - The Tianshu Tianji architecture is expected to surpass Blackwell in 2026 [1] - By 2027, the Tianshu Tianquan architecture aims to exceed NVIDIA's Rubin architecture [1] - After 2027, the company will shift focus to breakthrough computing chip architecture design [1] Group 2: Product Launch - Tianshu Zhixin has launched the "Tongyang" series of edge computing products [1] - The TY1000 has demonstrated performance exceeding NVIDIA's AGX Orin in tests related to computer vision, natural language processing, and DeepSeek 32B large models [1]

Core Insights - Domestic GPU companies are seizing the industry window period with intensive initiatives, exemplified by TianShu ZhiXin's release of its fourth-generation architecture roadmap, aiming to surpass NVIDIA's Rubin architecture by 2027 [1][2] Group 1: Product Development and Architecture - TianShu ZhiXin has introduced a roadmap for its four-generation GPU architecture, with specific timelines: "TianShu TianShu" in 2025 to surpass NVIDIA's Hopper architecture, "TianShu TianXuan" in 2026 to match Blackwell, and "TianShu TianQuan" in 2027 to exceed Rubin [3] - The company emphasizes energy efficiency and improved parallel computing resource utilization in its architecture design, achieving over 90% effective utilization in AI-related calculations [3][4] - The efficiency of the "TianShu TianShu" architecture is reported to be 60% higher than the industry average, with a 20% performance increase over NVIDIA's Hopper architecture in specific scenarios [4] Group 2: Product Launches and Market Positioning - TianShu ZhiXin has launched the "TongYang" series of products aimed at edge computing, which includes four models catering to various computational needs [5][6] - The "TongYang" series has demonstrated superior performance compared to NVIDIA's AGX Orin in multiple real-world scenarios, with the goal of becoming the leading domestic solution for edge computing [6][8] Group 3: Market Dynamics and Competitive Landscape - As of now, TianShu ZhiXin's products have served over 300 clients and completed more than 1,000 deployments across various sectors, indicating a growing commercial application [8] - The domestic GPU market is witnessing increased activity, with several companies like MoEr Thread, MuXi Co., and BiRan Technology completing IPOs, reflecting the ongoing interest in hard technology [8][9] - Despite the enthusiasm in the capital market, domestic GPU companies still hold a small market share compared to NVIDIA, which commands a significant global market presence [9]

Core Insights - The article discusses the significant advancements made by the Chinese AI computing company, Tensu Zhixin, which has recently launched a roadmap to surpass international giants like Nvidia's Hopper, Blackwell, and Rubin by 2025 to 2027 [2][4][5] - Tensu Zhixin's new architecture, the Tensu Tian Shu, has already demonstrated a performance improvement of approximately 20% over Nvidia's Hopper in key model scenarios, marking a substantial leap for domestic solutions [4][5][32] - The company aims to redefine the narrative of domestic GPU industry by moving away from a "benchmarking" approach to a self-defined leadership in AI computing [4][33] Group 1: Evolution of Computing in China - The roadmap released by Tensu Zhixin outlines a clear timeline and quantifiable breakthroughs, marking a departure from the traditional "follow and catch up" strategy adopted by many domestic companies [5][35] - The architecture is set to achieve a practical utilization rate of over 90% in executing attention mechanisms by 2025, showcasing the company's commitment to high efficiency [5][35] - The upcoming architectures, Tian Xuan and Tian Ji, are expected to further enhance performance and address industry-specific computational needs by 2026 [7][37] Group 2: Technological Innovations - Three core technological innovations underpin the aggressive roadmap: TPC BroadCast, Instruction Co-Exec, and Dynamic Warp Scheduling, which collectively enhance computational efficiency and resource utilization [10][39] - The company has adopted a problem-oriented research and development approach, addressing common industry pain points such as FP8 accumulation precision and matrix transposition overhead [11][40] - This focus on practical solutions has resulted in significant performance improvements, such as a 50% reduction in memory usage for model inference [14][44] Group 3: Redefining Value in Computing - Tensu Zhixin proposes a new value coordinate system for the computing industry, emphasizing high efficiency, predictability, and sustainability as key competitive factors [12][41] - The company aims to optimize total cost of ownership (TCO) by enhancing effective computing output per unit of power, thus reducing unnecessary costs for enterprises [14][43] - The design philosophy ensures that hardware remains adaptable to future algorithmic advancements, extending its lifecycle and value [16][44] Group 4: Market Positioning and Product Launch - The launch of the "Tongyang" series of edge computing products fills a gap in the domestic high-end edge computing market and establishes a comprehensive "cloud + edge + end" computing layout [18][46] - The product matrix covers a range of performance metrics from 100T to 300T, with specific models tailored for various applications [19][47] - Tensu Zhixin's products have already been deployed in over 300 customer environments, demonstrating their effectiveness in real-world applications [22][49] Group 5: Long-term Strategy in the GPU Industry - The competition in the GPU industry is fundamentally about building an open and collaborative ecosystem, which is essential for long-term success [52][52] - Tensu Zhixin is focused on creating a closed-loop ecosystem for domestic AI computing through hardware foundation, software adaptation, and partner collaboration [53][52] - The company aims to ensure that its products maintain value over a decade, positioning itself as a long-term player in the industry rather than merely a competitor to Nvidia [53][55]

如果在两年前问一家大模型公司最需要什么？答案是 "有没有卡"。但如果今天再问同样的问题，答 算力通胀 之下 都用不起了 "我们正在制造大量的垃圾算力。" 一位负责大模型训练集群的架构师曾这样抱怨。他的焦虑并不是没有风声。过去十年是算力野蛮增长 的十年，规模的快速扩张确实带来了阶段性的产业繁荣。但繁荣背后，是难以忽视的效率困局。 为了追赶 GPT-4 乃至 GPT-5 的能力，国内企业陷入了一场疯狂的参数竞赛。数以万计的 GPU 被 高度集成化塞进数据中心，它们日夜轰鸣，但产出的智能效益却并未如预期般线性增长。 这是一种典型的 "算力通胀"。行业习惯用芯片的理论峰值（ Peak Performance ）来衡量价值，但 在现实的复杂的训练任务中，这些昂贵的芯片往往"有力使不出"。 数据显示，在许多大规模训练集群中，算力 的有效利用率（ MFU ）仅能维持在 40% 左右，而在推 理场景下，大量的算力更是处于闲置状态，利用率甚至不足 20% 。 算法迭代与硬件僵化之间的错位也在加剧这种浪费，模型架构每六个月就发生一次巨变，从 案也许会变成"好不好用"。 天数智芯给出的架构路线图： 2025 年的天数天枢架构，超越 ...

中国GPU行业的发展，一直是悬在英伟达头上的达摩克利斯之剑。 过去的2025年，黄仁勋多次提及中国国产芯片企业正在迅速崛起，甚至成倍增长。同时还在多个场合警示，出口管制反而会加速中国企业的技术突围。 带来压迫感的向来不是某一个单一的对手，而是一片正在集体发力、依托全球最大应用市场快速迭代的产业生态。甚至越忌惮，有时就越不可说。 2026年1月24日，英伟达创始人黄仁勋如约在临近春节时现身上海办公室。虽然美国商务部刚刚有条件批准了英伟达向中国销售其最新一代AI芯片H200， 然而整个中国行，黄仁勋与员工们畅谈AI的过去未来，公司的目标愿景，却只字不提如何在中国销售H200。 房间里有头大象，但大家都假装没看见——英伟达对中国GPU日渐增长的忌惮，在这份沉默里，显得愈发具象化了一点。 忌惮来自于对结果的预知，悬而未决的只是时间——谁都知道，中国GPU早晚要踏上从"替代"到"超越"的征程，但什么时候开始，以前从来没人敢清晰地 说。 直到2026年开年，有中国GPU厂商用发布会，给出了一份自己的"答案"：2025年，天数天枢架构已经超越Hopper；2026年，天数天璇架构对标Blackwell； 2026年，天数天 ...

将投中网设为"星标⭐"，第一时间收获最新推送 作者丨 簪竹 来源丨 投中网 登陆港股18天后，天数智芯再次吸引众人目光。 1月26日，这家AI算力企业开了场发布会，不仅公布了四代架构路线图，同时还发布了边端算力产品"彤央"系列。 这家AI算力企业已经打通商业落地闭环，进入放量增长阶段。 从天数智芯透露的信息来看，他们2025年推出的天数天枢架构已超越英伟达目前主力的Hopper架构，在DeepSeek R3场景中实现性能超越Hooper 20%。此次发布会上，天数智芯正式公布了天枢之后的三代架构：2026年天璇对标Blackwell、天玑超越Blackwell，2027年天权超越英伟达目前最牛 的架构Rubin。 同样值得注意的是，天数智芯新发布的边端算力产品"彤央TY1000"，实测性能已超越英伟达AGX Orin。 随着"彤央"系列正式发布，再加上已推出 的"天垓"系列训练GPU和"智铠"系列推理GPU，标志着他们完成了"云+边+端"全场景算力布局。 通过互联网、科研、金融、医疗等领域众多客户1000多次实战部署，天数智芯的数千卡集群已稳定运行超过1000天。这也意味着，这家AI算力企业已 经打通商业落地 ...

Group 1 - The core viewpoint of the article is that TianShu ZhiXin has released a roadmap for its fourth-generation architecture, aiming to surpass NVIDIA's related products [2][3] - The roadmap includes the TianShu TianShu architecture expected to surpass Hopper by 2025, TianShu TianXuan architecture targeting Blackwell in 2026, TianShu TianJi architecture also surpassing Blackwell in 2026, and TianShu TianQuan architecture expected to surpass Rubin by 2027 [3] - The company plans to launch multiple products over the next three years based on this roadmap, continuously enhancing computing performance [3] Group 2 - Key details of the architectures include: TianShu TianShu architecture achieving over 90% effective utilization efficiency in AI calculations, TianShu TianXuan architecture introducing ixFP4 precision support, TianShu TianJi architecture covering all scenarios of AI and accelerated computing, and TianShu TianQuan architecture incorporating more precision support and innovative designs [3] - The company has reportedly achieved approximately 20% higher performance than the Hopper architecture in the DeepSeek V3 scenario due to core technology innovations [3] - TianShu ZhiXin was listed on the Hong Kong Stock Exchange on January 8, 2026, and its stock price closed at 188.2 HKD per share on January 26, 2026, reflecting a decline of 7.65%, with a market capitalization of 47.8 billion HKD [4]

【导读】天数智芯发布四代架构路线图 中国基金报记者 卢鸰 在港股上市不久的国产GPU公司天数智芯，发布了一份标注日期的、超越英伟达相关产品的 路线图。 2026年，天数天璇架构对标Blackwell； 2026年，天数天玑架构超越Blackwell； 2027年，天数天权架构超越Rubin； 2027年之后，将转向突破性计算芯片架构设计。 未来3年，天数智芯将基于此次发布的四代架构，陆续发布多款产品，持续提升计算性能。 在1月26日召开的天数智芯"智启芯程"合作伙伴大会上，天数智芯AI与加速计算技术负责人单 天逸公布了超越英伟达相关产品的四代架构路线图： 2025年，天数天枢架构超越Hopper； 天数智芯董事会主席盖鲁江同时兼任公司CEO，其于2020年7月加入公司，并于2025年5月 调任执行董事，主要负责公司的日常管理及整体策略规划。在此之前，盖鲁江拥有约17年财 务及投资经验。 2026年1月8日，天数智芯在港股上市。1月26日，天数智芯股价以188.2港元/股收盘，大跌 7.65%，最新市值为478亿港元。 编辑：江右 校对：王玥 制作： 鹿米 审核：陈墨 注：本文封面图由AI生成 单天逸还介绍了公 ...