明年超越英伟达Rubin？400亿国产GPU大消息

Core Viewpoint - TianShu ZhiXin has released a roadmap for its fourth-generation architecture, aiming to surpass NVIDIA's related products, as announced at the "Smart Start Chip Journey" partner conference on January 26 [2][4]. Group 1: Architecture Roadmap - In 2025, the TianShu TianShu architecture is expected to surpass NVIDIA's Hopper architecture [4]. - In 2026, the TianShu TianXuan architecture will benchmark against NVIDIA's Blackwell architecture [4]. - By 2026, the TianShu TianJi architecture is projected to exceed Blackwell [4]. - In 2027, the TianShu TianQuan architecture is anticipated to surpass Rubin [5]. - After 2027, the focus will shift to breakthrough computing chip architecture design [5]. Group 2: Product Development and Performance - Over the next three years, TianShu ZhiXin plans to release multiple products based on the announced fourth-generation architecture, continuously enhancing computing performance [6]. - The TianShu TianShu architecture supports high-precision scientific computing and AI precision computing, achieving over 90% effective utilization efficiency in attention mechanism-related calculations [7]. - The TianShu TianXuan architecture introduces ixFP4 precision support [7]. - The TianShu TianJi architecture aims for full-scene AI and accelerated computing coverage [7]. - The TianShu TianQuan architecture incorporates more precision support and innovative designs [7]. - Current core technology innovations have allowed TianShu ZhiXin to achieve approximately 20% higher performance than the Hopper architecture in the DeepSeek V3 scenario [7]. Group 3: Company Overview - The chairman and CEO of TianShu ZhiXin, Gai Lujian, has been with the company since July 2020 and is responsible for daily management and overall strategic planning [7]. - TianShu ZhiXin was listed on the Hong Kong Stock Exchange on January 8, 2026, and its stock closed at HKD 188.2 per share on January 26, experiencing a decline of 7.65%, with a market capitalization of HKD 47.8 billion [7].