Ko Jonghyun, Im Jiseong, Kim Jangsaeng, Shin Wonjun, Koo Ryun-Han, Park Sung-Ho, Woo Sung Yun, Lee Jong-Ho
Department of Electrical and Computer Engineering and Inter-university Semiconductor Research Center, Seoul National University, Seoul 08826, Republic of Korea.
Department of Electronic Engineering, Sogang University, Seoul 04107, Republic of Korea.
Sci Adv. 2025 Jul 18;11(29):eadt8227. doi: 10.1126/sciadv.adt8227.
In-memory computing (IMC) is a technology that enables efficient analog vector-matrix multiplication (VMM). This field has been extensively researched to overcome the performance bottlenecks associated with traditional von Neumann architectures. In addition to analog VMM, combining efficient neuromorphic modules with memory is essential to enable a broader range of IMC operations. Here, we propose a complementary metal-oxide semiconductor (CMOS)-compatible flash-gated thyristor-based neuromorphic module (FGTNM) that combines various functions in neural networks, such as quantization, nonlinear activation, and max pooling into a single module. The FGTNM features a small footprint (53 square micrometers) and low energy consumption (9.1 femtojoules per operation), outperforming previous CMOS-based modules. System-level IMC using the FGTNM shows a high accuracy (89.97%) on CIFAR-10 classification. This work showcases the potential to co-integrate various devices (flash memory, flash-gated thyristor, n-type metal-oxide semiconductor, and p-type metal-oxide semiconductor) on a single wafer, which broadens the scope of IMC applications beyond analog VMM.
内存计算（IMC）是一种能够实现高效模拟向量矩阵乘法（VMM）的技术。为克服与传统冯·诺依曼架构相关的性能瓶颈，该领域已得到广泛研究。除模拟VMM外，将高效的神经形态模块与内存相结合对于实现更广泛的IMC操作至关重要。在此，我们提出一种基于互补金属氧化物半导体（CMOS）兼容的闪存门控晶闸管的神经形态模块（FGTNM），它将神经网络中的各种功能，如量化、非线性激活和最大池化整合到一个模块中。FGTNM具有小尺寸（53平方微米）和低能耗（每次操作9.1飞焦）的特点，优于先前基于CMOS的模块。使用FGTNM的系统级IMC在CIFAR-10分类上显示出高精度（89.97%）。这项工作展示了在单个晶圆上共同集成各种器件（闪存、闪存门控晶闸管、n型金属氧化物半导体和p型金属氧化物半导体）的潜力，这拓宽了IMC应用的范围，使其超越模拟VMM。
