Core Insights - The National Supercomputing Internet Core Node has launched in Zhengzhou, marking the first operational deployment of a 30,000-card domestic AI computing pool in China [1] - The scaleX 10,000-card supercluster has achieved a significant milestone in AI computing, with a 20-fold increase in cabinet-level computing density and a 30% to 40% performance improvement for MoE trillion-parameter model training and inference scenarios [1][2] - The scaleX640 super node, which supports multi-brand acceleration cards and is compatible with mainstream computing ecosystems, aims to lower the R&D threshold for AI clusters and avoid redundant investments [2] Technological Advancements - The scaleX640 super node features an "open architecture" that enhances high-bandwidth, low-latency communication domains, enabling the formation of thousand-card computing units [1][2] - The supercomputing pool supports various large-scale AI computing scenarios, including trillion-parameter model training, high-throughput inference, and AI for Science applications, significantly improving research efficiency in fields like protein studies [2] Future Plans - The company plans to establish an Advanced Computing Technology Research Institute in Zhengzhou, with a 160,000 square meter R&D space and an initial goal of gathering 1,000 researchers focused on AI and high-performance computing applications [3] - The company aims to deepen technology research and application practices to provide more efficient, stable, and inclusive domestic intelligent computing services across various sectors of economic and social development [3]
国产AI算力里程碑时刻:中科曙光3套scaleX万卡超集群落地,最大国产AI算力池投入运营