魏少军呼吁：停用英伟达GPU

Core Viewpoint - The article emphasizes the need for China and other Asian countries to abandon reliance on NVIDIA GPUs for artificial intelligence training and inference, as this dependence poses long-term risks to regional autonomy and innovation [2][3]. Group 1: Call for Independence - Wei Shaojun, a prominent figure in China's semiconductor industry, advocates for the development of independent AI infrastructure in China, criticizing the current model that mimics the U.S. approach using NVIDIA and AMD GPUs [2][3]. - He warns that continued reliance on U.S. hardware could become "lethal" for the region's AI development, urging a strategic shift away from U.S. templates, particularly in algorithm design and computational infrastructure [2][3]. Group 2: Current Challenges - The U.S. government has imposed performance restrictions on AI and HPC processors that can be shipped to China, creating significant hardware bottlenecks and slowing down the training of advanced AI models [2]. - Despite these challenges, examples like the rise of DeepSeek demonstrate that Chinese companies can achieve significant algorithmic advancements without cutting-edge hardware [2]. Group 3: Future Directions - Wei suggests that China should focus on developing new types of processors specifically designed for training large language models, rather than continuing to rely on GPU architectures, which were originally intended for graphics processing [3]. - He acknowledges that while China's semiconductor industry has made progress, it still lags behind the U.S. and Taiwan, making it unlikely for Chinese companies to produce AI accelerators that rival NVIDIA's high-end products [3]. Group 4: NVIDIA's Dominance - NVIDIA GPUs dominate the AI field due to their large-scale parallel architecture, which is highly efficient for accelerating matrix-intensive operations in deep learning [4]. - The introduction of the CUDA software stack in 2006 allowed developers to write general code for GPUs, facilitating the standardization of deep learning frameworks like TensorFlow and PyTorch on NVIDIA hardware [4][5]. - Over time, NVIDIA has solidified its leading position through specialized hardware, tight software integration, and extensive cloud and OEM support, making its GPUs the default backbone for AI training and inference [5].