Improvement And Prediction On Reliability Of 3D NAND Flash

From the planar structure to the vertically stacked 3D structure,the storage density of NAND Flash continues to increase,which is a mainstream non-volatile memory.The superior storage performance and cheap manufacturing cost have made it widely used,and it plays an important role in the fields of solid-state disk,mobile storage and so on.3D NAND Flash is subject to a variety of reliability factors.On the one hand,the threshold voltage distribution is prone to change,especially in the process of data retention,resulting in more and more overlapping areas between different programming states.NAND Flash remains at a low RBER if the proper read offset is adopted.On the other hand,there is currently no mature failure prediction,and early failure warning of flash memory system is lacking.To improve the reliability of 3D NAND Flash,the data retention characteristics are studied,and a low-complexity mathematical model is designed to estimate the position of the optimal read voltage according to the deviation rule of read voltage.In the 64-layer 3D NAND Flash application case,the model reduces the RBER by 60+% compared with the default setting,and the gap from the theoretical optimal read is less than 10%.In more layers of 3D NAND Flash,whether it is a fully programmed block or a partially programmed block,the model has universality and room for improvement.On the other hand,this paper proposes a method for predicting block failure,which uses machine learning algorithms to monitor and predict the growth bad block.XGBoost algorithm performs best with an accuracy of 94.1% among three prediction algorithms.In addition,to explore the failure mechanism of 3D NAND Flash,quantified contribution of each feature to the prediction model has been discussed.