Precision-Sustained Analog In-Memory Computing using IGZO DRAM Array with Quasi-Nonvolatile and Continuously Tunable Weights

作者：C. Wang, X. Feng, Y. Shen, Z. Liao, S. Ding, Z. Li, X.-J. Yangdong, Y. Zhao, D. Kong, D.-C. Zhang, Y. Lv, M.-H. Xuan, X.-C. Chen, S.-J. Liang, F. Miao · 年份：2025 · DOI：10.1109/iedm50572.2025.11353772 · 研究领域：Advanced Memory and Neural Computing、Ferroelectric and Negative Capacitance Devices、Neural Networks and Reservoir Computing

We present a 58×32 compute-in-memory (CiM) array based on 4T1C IGZO DRAM cells, demonstrating precise vector-by-matrix multiplication (VMM) operations with quasi-nonvolatile and continuously tunable weights. Fabricated using a large-scale IGZO flexible panel process, the CiM array achieves ultra-long retention of 100,000 (or 2,000) seconds with 8-level (or 263-level) weight precision. The analog weights in CiM array can be precisely programmed into the target values, via a self-calibrating write scheme immune against device-level variation. Experimental results validate accurate VMM operations performed by our 58×32 IGZO CiM array, achieving a low root mean square error (RMSE) of 0.0078 μA between measured and expected results (48,000 data points). We also introduce a time-dependent ADC calibration strategy that maintains high VMM precision over extended periods without refresh.