MDC2025：全功能GPU路线清晰，MUSA生态进入规模化验证阶段

Investment Rating - The report does not explicitly state an investment rating for the industry or specific companies involved. Core Insights - The MUSA 5.0 has established a comprehensive full-stack system that includes instruction sets, programming models, compilers, and communication libraries, achieving engineering performance close to international mainstream standards across key metrics [2][10] - The Huagang architecture, introduced at the MDC 2025, represents a significant upgrade in compute density, energy efficiency, precision coverage, and interconnect capabilities, supporting full-precision computing from FP4 to FP64 and introducing mixed low precision [2][10] - Moore Threads is one of the few domestic GPU vendors committed to a "full-function GPU" strategy rather than focusing solely on AI accelerators, indicating a long-term vision for broader ecosystem development [2][10] Summary by Sections Event Overview - The inaugural MUSA Developer Conference (MDC 2025) was held on December 20-21, 2025, in Beijing, focusing on sovereign computing and the developer ecosystem, unveiling the next-generation full-function GPU architecture Huagang and the Ku'e ten-thousand-card AI compute cluster [1][9] Technical Developments - The Huagang architecture emphasizes asynchronous programming and ultra-large-scale interconnect (MTLink), laying the groundwork for scaling to ten-thousand-card and hundred-thousand-card clusters [2][10] - The Ku'e ten-thousand-card AI compute cluster achieved approximately 60% MFU on dense models and 40% on MoE models, with a linear scaling efficiency of about 95% and effective training time exceeding 90% [3][11] Ecosystem and Strategy - The report outlines a clear roadmap for progressively open-sourcing core components, including compute libraries and communication libraries, to enhance the ecosystem [2][14] - The MT Lambda platform was launched, integrating physics engines, graphics rendering engines, and AI compute engines to create a full-stack framework for development, simulation, and training [3][12] Future Directions - The company has articulated a clear product segmentation path with a focus on unified AI training and inference, positioning itself as a foundation for next-generation AI factories [2][14] - The Huashan and Lushan architectures are designed to cater to AI training and high-performance graphics rendering, respectively, with significant improvements in various performance metrics [3][14]