부산대학교 도서관

Although non-volatile memory (NVM) based analog computation-in-memory (ACiM) may be the potential of highly energy efficient AI accelerators, it has been challenging to realize the real benefit due to non-ideal characteristics of ReRAM. In this study, a scalable 256kb 1T1R array-based wafer-level holistic co-optimization was demonstrated for the improvement of ReRAM based ACiM characteristics. We developed a methodology to enhance retention properties, resulting in a 65% reduction in weight variation and a 49% improvement in relaxation. With the realistic retention results, we could achieve a 124% improvement in inference accuracy on CIFAR-10 using a holistic approach. In addition, we propose a comprehensive mitigation method for IR drop for large ReRAM arrays.