北京发布升级版“系统软件栈”，支持全球20多款主流AI芯片

Core Insights - The 2025 Artificial Intelligence Computing Conference was held in Beijing, where the upgraded "system software stack" called "Zhongzhi FlagOS" v1.5 was released, marking a new phase in the global AI underlying technology ecosystem focused on "open computing" [1] - The development of large AI models requires significant computing power, and the differences in software ecosystems and interfaces among various chips complicate model migration and adaptation, limiting the application of diverse computing resources [1] Group 1 - The "Zhongzhi FlagOS" aims to provide unified, open, and efficient computing support for AI models, addressing compatibility issues among different AI chips [1] - The system software layer enhances computing resource utilization, simplifies model migration, and supports multiple intelligent application scenarios, becoming a critical infrastructure for supporting large models in the scale of billions [1] - The "Zhongzhi FlagOS" v1.5 has achieved breakthroughs in four dimensions: comprehensiveness, automation, performance, and application scenarios [2] Group 2 - The upgraded system fully supports the development and deployment of robot "brain" and "small brain" models, providing robust system support for smarter robots [2] - Improvements in parallel strategies and scheduling algorithms have led to a maximum training acceleration of 36.8% and a 20% acceleration in inference for typical large model tasks [2] - The "Zhongzhi FlagOS" v1.5 supports over 20 mainstream AI chips globally, and the general operator library FlagGems has become the largest Triton language operator library, with over 200 core operators [2] Group 3 - To promote AI system technology talent cultivation, the Beijing Academy of Artificial Intelligence has launched the "Zhongzhi FlagOS" university program in collaboration with over 10 universities, including Peking University and Tsinghua University [3] - The "Super Node Intelligent Computing Application 'Beijing Plan'" was officially released, focusing on creating an inclusive AI computing foundation for various scenarios such as embodied intelligence, intelligent manufacturing, and smart healthcare [3] - The Beijing Artificial Intelligence Standardization Technical Committee was established, along with the "Beijing Reconfigurable Computing Soft and Hardware Collaborative Technology Innovation Center" [3]