21专访｜中昊芯英CTO郑瀚寻：国产AI芯片也将兼容不同平台

Core Insights - The demand for AI computing is driving attention towards non-GPU AI chips, with companies like Google and Groq leading the way in alternative architectures [1][2] - The rise of custom ASIC chips is notable, as companies seek to reduce costs and enhance personalized AI capabilities [1][2] - The trend of exploring opportunities beyond GPU chips is becoming increasingly evident in the market [1] Market Trends - New players in Silicon Valley, such as Groq and SambaNova, are focusing on architectural innovation rather than GPU-like structures to achieve performance breakthroughs [2] - The success of GPU chips is largely attributed to NVIDIA's established engineering teams, making it challenging for new entrants to replicate this success [2] - Custom ASIC chips are gaining traction, as evidenced by Broadcom's significant orders and Google's ongoing development of TPU chips [2] Technological Developments - The investment in Tensor Processing Units (TPUs) is seen as cost-effective, especially in the era of large models, where data transmission scales significantly enhance computational efficiency [3][4] - TPUs are compared to 3D printers in their ability to efficiently handle computation tasks, leading to better data migration and lower energy consumption [4] - The challenge for domestic XPU chips lies in scaling "single-point efficiency" to "cluster efficiency" to meet the demands of large-scale AI computing [4][5] Infrastructure and Connectivity - Future data transmission is identified as a potential bottleneck for AI infrastructure, with Tensor Cores offering advantages in handling increased data volumes [5] - Middle and high-speed interconnect capabilities are being developed, with companies like 中昊芯英 supporting large-scale chip interconnectivity [5][6] - The evolution of Ethernet technology has made it competitive for AI chip manufacturers, with significant improvements in physical media and bandwidth capabilities [6] Software Ecosystem - The development of a robust software ecosystem is crucial, as domestic chip platforms must build their own software stacks to ensure compatibility and performance [6][7] - The ongoing evolution of large language models, primarily based on the Transformer architecture, presents opportunities for AI chip manufacturers to align their product development with these advancements [7]