Process optimization and validation of the polysilicon grain-boundary barrier for high-accuracy 3-dimensional NAND compute-in-memory chips

Read full text Summarise with VIRAT
Authors: ZHENG Hao, LIU Huiwen, XU Kezhi, Zhang Baotong, YANG Yuancheng, XIA Zhiliang, HUO Zongliang
Cited by: 0
References: 26

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

随着人工智能与边缘计算的快速发展, 存算一体(computing-in-memory, CIM)架构被认为是缓解冯·诺依曼瓶颈的重要技术路径. 基于3维(3D) NAND的CIM方案兼具高存储密度与工艺成熟度, 但在执行矩阵-向量乘法(matrix-vector multiplication, MVM)等模拟计算任务时, 串电流分布展宽会导致累加电流偏差, 进而引起计算精度下降问题. 其中3D NAND顶部选择(top select gate, TSG)晶体管沟道的多晶硅晶界更是会直接影响到串电流的分布. 因此本文采用计算机辅助设计软件(technology computer-aided design, TCAD)建立TSG Deck器件模型, 分析TSG多晶硅沟道晶界陷阱诱发的势垒对开态电流波动的影响规律. 在此基础上, 提出一种通过多晶硅前驱体组合调控实现等效氢钝化窗口优化的工艺方案, 并对不同工艺分组条件下的开态电流分布进行单片晶圆尺度的统计评估. 结果表明, 最优工艺可使位线端电流分布的归一化标准差较优化前减小50%, 并在CIM系统级仿真中使GPT-2 124M模型的INT8推理中的MVM计算误差相对基准工艺降低14.7%—66.8%. 综上, 本工作为面向高精度3D NAND CIM芯片的工艺优化方案提供了可实现的设计依据.

Fields of Study & Subjects

Advanced Memory and Neural Computing Advancements in Semiconductor Devices and Circuit Design Semiconductor materials and devices

Publication details

Journal / source
Acta Physica Sinica
Volume
75
Issue
8
Publication date
1 January 2026
Language
ZH