PENN: Predicting Errors Using Neural Network in Phase Change Memory

Apoorva Rumale, Jake Ekoniak, Marjan Asadinia

Abstract: Phase Change Memory (PCM) is an emerging non-volatile memory technology with the potential to replace DRAM as main memory, thanks to its scalability and high storage capacity. However, increasing the density of PCM—particularly in Multi-Level Cell (MLC) configurations—raises its vulnerability to errors. To address these challenges, we studied data related to write patterns, cell states, and electrical input parameters to identify complex correlations between these factors and the occurrence of errors. Leveraging these insights, we developed PENN, a predictive model built with a multi-layer perceptron neural network to accurately forecast error types. By utilizing Deep Neural Networks, PENN uncovers hidden patterns within the data, improving error prediction and facilitating more effective error correction strategies. The experimental findings demonstrate that our model, including the multi-layer perceptron regression, achieved notable performance. In the case of Write Disturbance Error prediction, we obtained a Mean Squared Error (MSE) of 8.36 × 10^-6, a Mean Absolute Error (MAE) of 0.0022, and a Coefficient of Determination (R^2) of 0.997. For Bit Flip error prediction, the MSE was 7.44 × 10^-5 , the MAE was 0.0045, and the value reached 0.957.