During the 2025 SOHU Technology Annual Forum held on May 17, Zheng Weimin, a professor in the Department of Computer Science at Tsinghua University and an academician of the Chinese Academy of Engineering, emphasized the necessity of diversifying large model training systems. Currently, the industry primarily relies on two systems: NVIDIA GPUs and domestic chips. While NVIDIA GPUs excel in hardware performance and ecosystem maturity, they are hampered by challenges such as sales bans, skyrocketing prices, and supply shortages. On the other hand, despite investments from over 30,000 domestic enterprises in R&D, and gradual improvements in hardware performance, domestic chips continue to grapple with a crucial obstacle: insufficient ecosystem compatibility.

In response to these challenges, Zheng Weimin proposed two strategic breakthroughs: Firstly, the establishment of a CUDA-analog system to minimize the learning curve for developers and foster a new ecosystem. Secondly, enhancing the hardware capabilities of domestic chips to exceed 60% of comparable foreign products, ideally aiming for 70%-80%, while simultaneously optimizing the ecosystem to attract and retain users.