直击WAIC丨如何缓解AI训练“效率瓶颈”？摩尔线程张建中：打造AGI“超级工厂”

Core Insights - The 2025 World Artificial Intelligence Conference (WAIC 2025) is being held in Shanghai from July 26 to 28, where the concept of "AI Factory" was introduced by Moore Threads [1][3] - The CEO of Moore Threads, Zhang Jianzhong, emphasized the need for innovative engineering solutions to address the efficiency bottlenecks in large model training due to the explosive growth of generative AI [1][3] Group 1: AI Factory Concept - The "AI Factory" is likened to the process upgrades in chip wafer fabs, requiring innovations in chip architecture, overall cluster architecture optimization, software algorithm tuning, and resource scheduling system upgrades [3] - The efficiency of the AI Factory is determined by five core elements, summarized in the formula: AI Factory Production Efficiency = Accelerated Computing Generality × Single Chip Effective Computing Power × Single Node Efficiency × Cluster Efficiency × Cluster Stability [3] Group 2: Technological Innovations - Moore Threads' GPU single chip, based on the MUSA architecture, integrates AI computing acceleration, graphics rendering, physical simulation, and ultra-high-definition video encoding capabilities, supporting a full precision spectrum from FP64 to INT8 [3] - The use of FP8 mixed precision technology in mainstream large model training has resulted in a performance increase of 20% to 30% [3] Group 3: Memory and Communication Efficiency - The memory system of Moore Threads achieves a 50% bandwidth saving and a 60% reduction in latency through various technologies, including multi-precision near-memory reduction engines and low-latency Scale-Up [4] - The ACE asynchronous communication engine reduces computational resource loss by 15%, while the MTLink 2.0 interconnect technology provides 60% higher bandwidth than the domestic industry average, laying a solid foundation for large-scale cluster deployment [4] Group 4: Reliability and Fault Tolerance - The introduction of zero-interruption fault tolerance technology allows for the isolation of affected node groups during hardware failures, enabling uninterrupted training for the remaining nodes [4] - This innovation results in an effective training time ratio exceeding 99% for the KUAE cluster, significantly reducing recovery costs [4]