Core Viewpoint - The article highlights the launch of the "Huagang" GPU architecture by Moer Thread, showcasing its advancements in AI training and high-performance graphics rendering, as well as the company's commitment to developing a comprehensive computing ecosystem based on its MUSA architecture [1][2][11]. Group 1: Product Launches and Innovations - Moer Thread introduced the "Huagang" architecture, which features a new instruction set that enhances computing density by 50% and improves efficiency by 10 times under the same manufacturing conditions [2]. - The "Huashan" chip, designed for AI training and inference, offers floating-point computing power, memory bandwidth, and high-speed interconnect capabilities that are competitive with NVIDIA's Blackwell and Hopper architectures [4][5]. - The "Lushan" chip focuses on high-performance graphics rendering, achieving a 64-fold increase in AI computing performance and a 50% improvement in ray tracing performance [5]. Group 2: Infrastructure and Ecosystem Development - The "Kua'e" supercomputing cluster, based on the "Huagang" architecture, boasts a floating-point computing capability of 10 Exa-Flops, with a 60% utilization rate for Dense models and 40% for MoE models [5][8]. - Moer Thread upgraded its MUSA architecture to version 5.0, achieving over 98% efficiency in core computing libraries and planning to open-source key components [8]. - The company is also developing a MUSA ecosystem center to support research and innovation in cutting-edge fields such as quantum technology and AI for 6G [10]. Group 3: Market Position and Financial Performance - Moer Thread, founded in October 2020, has seen significant market interest, with its stock price increasing by 481% since its debut on the STAR Market at 114.28 CNY per share, reaching approximately 664.10 CNY per share and a market capitalization of about 312.15 billion CNY [11].

（原标题：摩尔线程发布"花港"架构，力推国产万卡智算生态） 21世纪经济报道记者/彭新 在登陆科创板近半月之际，12月20日，摩尔线程（688795.SH）于北京举办首届MUSA开发者大会。 会上，摩尔线程展示了其基于MUSA统一架构为核心的全栈技术成果，并发布下一代全功能GPU架 构"花港"。 "华山"芯片主要面向AI训推一体与超大规模智能计算，可为智算集群提供算力支撑，张建中称该芯片的 浮点算力、访存带宽、高速互联带宽三个指标的能力介于英伟达量产的Blackwell架构和上一代Hopper 架构之间，而在访存容量指标上优于上述两个架构。 同步发布的"庐山"芯片则侧重于高性能图形渲染，其AI计算性能提升64倍，几何处理性能提升16倍，光 线追踪性能提升50倍，并增强了纹理填充和显存容量。 基于该架构，摩尔线程推出面向AI训推一体的"华山"芯片及高性能图形渲染的"庐山"芯片，并展示 了"夸娥"万卡智算集群技术成果与搭载"长江"SoC芯片的端侧AI产品。 MUSA（Meta-computing Unified System Architecture）为摩尔线程自主研发的"元计算"统一计算架构， 覆盖从芯片架构、 ...

摩尔线程发布了全新的长江智能SoC，基于该芯片推出了MTT AIBOOK笔记本电脑，32GB+1TB售价9999元，将于2026年1月10日正式开售。AIBOOK支持MUSA统一架构，异构AI算力高达50TOPS，支持32GB LPDDR5X 7500MT/s统一内存。 https://t.co/4QJbGyj5BW ...