Core Insights - The article discusses a significant advancement in AI hardware through a high-precision analog in-memory computing chip developed by Nanjing University, which has achieved the highest precision record in the field [1][2]. Group 1: Research and Development - The research team at Nanjing University has proposed a high-precision analog in-memory computing scheme and developed a chip based on complementary metal-oxide-semiconductor (CMOS) technology [1]. - The chip utilizes a weight remapping technique to enhance computational accuracy, achieving a root mean square error (RMSE) of only 0.101% in parallel vector-matrix multiplication operations, setting a new record for precision in analog vector-matrix multiplication [2]. Group 2: Performance in Extreme Conditions - The chip demonstrates stable operation in extreme environments, maintaining RMSE levels of 0.155% at -78.5℃ and 0.130% at 180℃ [2]. - Measurements in strong magnetic fields showed that the output current of the chip's core unit varied by no more than 0.21% compared to conditions without a magnetic field, confirming the reliability of the high-precision analog computing scheme under extreme conditions [2]. Group 3: Implications for AI Hardware - This breakthrough is seen as a critical step towards the practical application of analog in-memory computing technology, with the potential to promote the development of low-power, high-precision AI hardware [2].

科技日报记者 金凤 记者10日从南京大学获悉，该校类脑智能科技研究中心研究团队提出了一种高精度模拟存内计算方案， 并以此为基础，研发出一款基于互补金属氧化物半导体工艺的模拟存算一体芯片。相关成果近日刊发于 国际学术期刊《科学进展》。 结合权值重映射技术，该芯片在并行向量矩阵乘法运算中实现了仅0.101%的均方根误差，创下了模拟 向量-矩阵乘法运算精度的最高纪录。 来源：科技日报 论文共同通讯作者、南京大学教授梁世军介绍，该芯片在-78.5℃和180℃的极端环境下依然能稳定运 行，矩阵计算的均方根误差分别维持在0.155%和0.130%的水平。 测试结果证实了高精度模拟计算方案在极端环境下的可靠性，这项突破有望推动低功耗、高精度AI硬 件技术的落地。 ...