Research On Error Model Of 3D QLC NAND Flash

Our country is vigorously promoting the independent development of semiconductor technologies,and NAND flash memory is one of the key concerns and investment objects.3D QLC NAND Flash is the highest packing density and lowest per-bit cost NAND Flash available in market.Solid State Drive using QLC NAND Flash delivers a per-bit cost close to an ordinary Hard Disk Drive,and at the same time an overwhelming power consumption advance.However,as a data storage component,NAND flash memory chip has an intrinsic limited lifecycle problem.Compared with the common TLC flash memory in today's market,the read voltage window between adjacent voltage states of QLC flash memory is narrower,the threshold voltage easily overlaps more areas,and the data reliability risk is higher.Therefore,aiming at the high reliability risk of QLC flash memory,this dissertation uses 96layers 3D QLC flash memory which is more advanced in industry to carry out reliability research and build error model.First,flash memory error is closely related to the distribution of threshold voltage.In order to analyze the influence of various risk factors on the distribution of threshold voltage,this dissertation first obtains the image of threshold voltage distribution under various test conditions,and analyzes the characteristics of the change of threshold voltage distribution caused by various risk factors;Second,based on the analysis of the risk factors,this dissertation divides the Read Disturb and Data Retention into two groups of experiments.Then we build the test platform and make the test plan,use customized Windows test driver and python script program to collect and analyze the original error bit data of flash memory chip under various experimental conditions;Third,based on the analysis of the test data,it is noticed that the error characteristics of different Word Line and logical pages are distinctive.In this dissertation,the traditional function fitting method and three machine learning methods are used to build error models of different word lines and logical pages,and then multiple indicators are used to evaluate.The evaluation result shows that using function fitting method,the average RBER of all Word Line and four logical page type mostly reach a Coefficient of Determination(R~2)around 0.95;and Random Forest shows the best error bits prediction accuracy among three machine learning methods.