【2025数博会】华为这样解决单芯片算力差距短板

Group 1 - The core event was the launch of Huawei's CloudMatrix 384 super node cluster during the 2025 Digital Expo, emphasizing the theme "Computing Power Gathering, Intelligent Future" [1] - Huawei's CTO explained the transition from small models to large models, highlighting the need for deeper thinking and multi-modal capabilities, driven by market demand [3] - The CloudMatrix 384 architecture consists of 16 cabinets, with 12 for computing power and 4 for networking, achieving a latency of approximately 150 nanoseconds between chips [3] Group 2 - The CloudMatrix 384 super node cluster can achieve a computing power scale of up to 300P, surpassing industry standards by 67%, and improving training efficiency by over 41% [4] - The architecture is designed to lower model training costs significantly, addressing current market needs [4] - Huawei's 384 nodes exceed NVIDIA's offering of 72 nodes, positioning Huawei's solution as 1.67 times more powerful while maintaining lower bandwidth latency [3]