Core Insights - The first MUSA Developer Conference (MDC2025) was held by Moore Threads in Beijing, emphasizing the advanced and leading technology of the MUSA architecture as the foundation for full-function GPUs [1] - The company aims to enhance its research and development investments to tackle core technological challenges from hardware to software, focusing on building a self-reliant domestic computing industry ecosystem [1] Group 1 - The MUSA architecture supports full precision computing from FP4 to FP64, with a 50% increase in density and a 10-fold improvement in efficiency, capable of supporting computing clusters exceeding 100,000 cards [1] - Moore Threads plans to launch two high-performance chips based on the MUSA architecture: the "Hua Shan" chip for AI training and inference, and the "Lu Shan" chip for high-performance graphics rendering, both set to be mass-produced next year [1] Group 2 - The company showcased the "Kua E Wan Ka" intelligent computing cluster, demonstrating its capability to support trillion-parameter model training with reliability and achieving international mainstream levels in several key precision metrics [2] - Moore Threads shared its planning for the MTT C256 super-node architecture aimed at the next generation of ultra-large-scale intelligent computing centers [2] - The company also released the MTTAIBOOK, an AI computing device equipped with the intelligent SoC chip "Chang Jiang" [2]
摩尔线程发布新一代GPU架构“花港” 支持十万卡集群