摘要
Spiking Neural Networks (SNNs) implemented with Resistive RAM (ReRAM) offer promising advantages in area and power efficiency due to their compatibility with compute-in-memory architectures. However, process-induced resistance variability in ReRAM cells poses a significant challenge to inference accuracy. To address this issue, we propose a test and calibration framework to maintain target model accuracy. The test flow employs systematic pattern generation and formulates a set of linear equations to estimate ReRAM cell resistances. Given the estimated resistance values, rows exhibiting large deviations are replaced using redundant rows to mitigate computational errors. Experimental results on Tiny ImageNet with a Transformer-based SNN demonstrate that the proposed calibration method improves inference accuracy by 2.3% and 4.2% with one and two redundant rows, respectively.