RRAM-Based Isotropic CNNs with High Robustness and Resource Utilization Rate
Wenyong Zhou, Yuan Ren, Jiajun Zhou, Chenchen Ding, Zhengwu Liu, Ngai Wong · 2025
Resistive random-access memory (RRAM)-based compute-in-memory (CIM) systems show great potential for accelerating convolutional neural networks (CNNs). However, classical RRAM-based CNNs suffer from performance degradation from weight quantization and device non-idealities, and resource under-utilization due to mismatches between weight matrices and crossbar arrays. In this work, we propose RRAM-based isotropic CNNs that enhance model robustness and improve resource utilization concurrently. Extensive simulations demonstrate that the isotropic CNNs yield up to 3.74% and 12.61% accuracy improvement under quantization and non-idealities, respectively. Moreover, they increase the utilization rate by 5.2-13.4% in typical network architectures. These results make our design a highly robust and efficient RRAM-based CNN solution.