ML-PreP: Machine Learning Based Error Prediction for Phase Change Memory
Jake Ekoniak, Apoorva Rumale, Marjan Asadinia
Abstract
Phase Change Memory (PCM) is a non-volatile memory technology that shows great promise as a potential replacement for DRAM in main memory due to its scalability, low read latency, and potential for high storage capacity. However, as PCM-specifically Multi-Level Cell (MLC) PCM-becomes denser, it becomes increasingly susceptible to errors, a problem that current error correction strategies cannot efficiently address. In response, we have developed a predictive model, ML-PreP, specifically tailored for MLC PCM. This model is trained on MLC PCM data, including write patterns, cell states, and electrical input parameters, to understand the complex relationships between these inputs, output parameters, and error occurrences. Our model employs a multilayer perceptron network and an AdaBoost regression model to demonstrate high accuracy in forecasting error types. Subsequently, a convolutional neural network estimates the number of errors per line, while an additional detection model pinpoints their locations, enabling the efficient application of standard error correction mechanisms. This approach underscores the potential of machine learning-driven error prediction to significantly enhance the robustness and efficiency of MLC PCM systems.