Yangdong Xing-Jian, Wang Cong, Zhao Yichen, Wang Zi-Chun, Yang Zaizheng, Liu Zenglin, Yu Wentao, Zeng Zhoujie, Wang Shuang, Wei Wei, Shen Yu, Kong Dehe, Ding Shuo, Wang Xu, Pan Chen, Liang Shi-Jun, Miao Feng
Institute of Brain-Inspired Intelligence, National Laboratory of Solid-State Microstructures, School of Physics, Collaborative Innovation Center of Advanced Microstructures, Nanjing University, Nanjing, China.
Institute of Interdisciplinary Physical Sciences, School of Physics, Nanjing University of Science and Technology, Nanjing, China.
Sci Adv. 2025 Sep 12;11(37):eady4798. doi: 10.1126/sciadv.ady4798.
Analog computing has gained increasing attention for its potential in artificial intelligence hardware. The computation in traditional analog systems relies on use of intrinsic physical quantities (e.g., resistance), which are prone to fluctuations due to environmental changes or repeated programming, leading to compromised precision. Here, we shift the reliance on intrinsic physical quantity of memory devices to geometric ratio of transistors, enabling ultrahigh-precision analog computation. We demonstrate an analog in-memory computing chip based on a standard complementary metal-oxide semiconductor process, achieving the highest precision reported to date. Enhanced by the proposed weight remapping technique, the chip realizes ultrahigh computing accuracy with a root mean square error of only 0.101% across multiple parallel vector-by-matrix multiplication operations. Moreover, our analog in-memory computing chip maintains high precision, with an error of 0.155 and 0.130% under environmental temperatures of -78.5° and 180°C, respectively. This work pushes the boundaries of analog computing precision by leveraging stable geometry feature of devices.
模拟计算因其在人工智能硬件方面的潜力而受到越来越多的关注。传统模拟系统中的计算依赖于固有物理量（如电阻）的使用，这些物理量由于环境变化或重复编程而容易出现波动，从而导致精度受损。在此，我们将对存储器件固有物理量的依赖转变为对晶体管几何比例的依赖，实现了超高精度的模拟计算。我们展示了一款基于标准互补金属氧化物半导体工艺的模拟内存计算芯片，达到了迄今为止报道的最高精度。通过所提出的权重重映射技术的增强，该芯片在多个并行向量与矩阵乘法运算中实现了超高的计算精度，均方根误差仅为0.101%。此外，我们的模拟内存计算芯片保持了高精度，在环境温度为-78.5°和180°C时，误差分别为0.155%和0.130%。这项工作通过利用器件稳定的几何特征拓展了模拟计算精度的边界。
