ScholarWorks@성균관대학교

초록

"Compute-in-memory (CIM) technology based on emerging nonvolatile memories (NVMs) has shown promise in enhancing artificial intelligence applications by integrating computation directly within NVM arrays. However, the efficiency of CIM systems is often curtailed by the substantial overhead that is caused by traditional complementary metal-oxide-semiconductor (CMOS)-based analog-to-digital converters (ADCs). Here, we report an in-memory ADC (IMADC) that leverages NVMs to perform the dual functionalities of reference generation and voltage comparison, effectively minimizing the area occupancy and energy consumption, is reported. The IMADC not only significantly outperforms traditional ADCs but also enables the inherent processing of nonlinear activation functions such as the sigmoid function, which is required for neural networks. The IMADC-based CIM system achieves software-comparable accuracy in CIFAR-10 image classification on the VGG-9 network. The IMADC exhibits significantly reduced area occupancy (45 μm2) and energy consumption (29.6 fJ) compared to conventional CMOS-based ADCs. The IMADC, compatible with various types of NVMs, demonstrates significant potential for enhancing the efficiency of CIM systems in terms of area occupancy and energy consumption. © 2024 The Author(s). Advanced Intelligent Systems published by Wiley-VCH GmbH.

키워드