PDA-KNN: Page data arrangement based KNN to reduce retention errors for QLC 3D NAND flash memories

<p dir="ltr">Multi-bit cell storage technology is one of the main way to expand the storage capacity of 3D NAND flash memories. Quarter level cell (QLC) 3D NAND flash stores 4 bit per cell, resulting in narrower read margins. Constrained to read margins, QLC always takes threshold voltage drifting caused by lateral charge migration (LCM) and vertical charge loss (VCL), which causes a large retention error. To suppress charge leakage effect, there are some algorithm in form of intra-page data mapping such as WBVM, DVDS, etc. However, we observe page data distribution and pattern also contribute to the impacts of LCM and VCL. Thus, we proposed PDA-KNN to arrange page data for LCM and VCL suppression. PDA-KNN applies a K-Nearest-Neighbout (KNN) clustering make each page level data similar. The clustering is to minimize the global impacts derived from the LCM among wordlines. Since four clustered data classes should reflection to a explicit storage page level in QLC, we design a loss function with Bayesian principle and Gray code, which is to reshape voltage threshold pattern to reduce the VCL and LCM. Finally, we arrange data based on the Hamming Distance to ensure adjacent patterns are as similar as possible, to further decrease the LCM. Thereby the arranged data alleviate both LCM and VCL, decreasing the bit error rate (BER) during data retention. In addition, PDA-KNN do not need extra flag bits to record data reshape compared with WBVM, DVDS. The experiment on five types of real world data results show that the PDA-KNN reduces the average BER by 85.4$\%$ compared with strategy without data arrangement, and by 20.5$\%$, 18.8$\%$ compared respectively with WBVM and DVDS with code-length 64. </p>