Resistive random-access memory (RRAM)-based computing-in-memory (CIM) has emerged as a promising solution for energy-efficient deep learning inference by eliminating the memory bottleneck of conventional von Neumann architectures. However, the limited bit precision and inherent variability of memristive devices pose significant challenges to inference accuracy. In this study, we present a post-training quantization (PTQ) framework that incorporates σ-based clipping to suppress weight outliers before quantization. This technique narrows the dynamic range of weights, improving quantization resolution and robustness. Using a three-layer fully connected neural network trained on binarized MNIST data, we analyzed the impact of σ-clipping on PTQ performance across z-score thresholds. Clipping improved inference accuracy and reduced accuracy variance across retrained models. Furthermore, we experimentally validated our approach using a 48 × 48 Al2O3/TiOx-based RRAM crossbar array. Despite conductance deviations during weight programming, clipping improved inference accuracy from 96.27% to 97.76%, confirming its effectiveness under hardware variation. In simulation, when variation corresponding to experimentally observed fluctuation was introduced, PTQ accuracy dropped from 96.42% to 83.67%, while the clipped model maintained 96.96%, demonstrating a 12.75%p improvement. These results establish σ-clipping as a practical and lightweight strategy to enhance the quantization robustness of CIM-based neural networks.
Oh et al. (Mon,) studied this question.