Efficient Calibration for RRAM-based In-Memory Computing using DoRA
This addresses the problem of maintaining accuracy in edge AI systems with RRAM, offering an energy-efficient solution for real-world deployment.
The paper tackles accuracy degradation in RRAM-based in-memory computing due to conductance drift by proposing a DoRA-based calibration framework that restores accuracy without modifying RRAM weights. Experiments on ResNet50 show 69.53% accuracy restoration using only 10 calibration samples and updating 2.34% of parameters.
Resistive In-Memory Computing (RIMC) offers ultra-efficient computation for edge AI but faces accuracy degradation due to RRAM conductance drift over time. Traditional retraining methods are limited by RRAM's high energy consumption, write latency, and endurance constraints. We propose a DoRA-based calibration framework that restores accuracy by compensating influential weights with minimal calibration parameters stored in SRAM, leaving RRAM weights untouched. This eliminates in-field RRAM writes, ensuring energy-efficient, fast, and reliable calibration. Experiments on RIMC-based ResNet50 (ImageNet-1K) demonstrate 69.53% accuracy restoration using just 10 calibration samples while updating only 2.34% of parameters.