绕开光刻机“卡脖子” 中国新型芯片问世！专访北大孙仲：支撑AI训练和具身智能 可在28纳米及以上成熟工艺量产

Core Viewpoint - A Chinese research team has developed a novel high-precision, scalable analog matrix computing chip based on resistive random-access memory (RRAM), achieving 24-bit fixed-point precision for the first time globally, which allows for reduced computational card usage for similar tasks [1][12]. Group 1: Technology and Innovation - The new chip represents a significant leap in precision, reducing the relative error from 1% to one part in ten million (10^-7), thus addressing the historical precision bottleneck of analog computing [7][15]. - The chip can support advanced applications such as 6G, embodied intelligence, and AI model training, and can be produced using mature processes of 28nm and above, circumventing the bottleneck of photolithography [1][12]. - The research team has introduced a new feedback circuit and utilized classic iterative optimization algorithms to enhance precision without sacrificing energy efficiency or speed [11][15]. Group 2: Market and Application Potential - The chip is particularly suited for applications requiring large-scale matrix operations, such as AI model training, 6G communications, and supercomputing tasks, which are fundamentally based on matrix calculations [10][20]. - The team aims to scale the matrix size from 16x16 to 128x128 within two years, with a long-term goal of reaching 512x512, which would enable practical applications in medium-scale scenarios [24][25]. - The technology provides a potential alternative to reliance on advanced processes and NVIDIA GPUs, positioning the team at the forefront of modern analog computing [10][26]. Group 3: Strategic Importance - This development offers a "detour" for China's computing capabilities, potentially reducing dependence on advanced manufacturing processes and foreign technology [10][26]. - The successful demonstration of this new path confirms its potential, although significant investment and collaboration across the technology and industry sectors will be necessary to fully realize its capabilities [10][26].