Approximate 3D-TLC NAND Flash Write with Initial Error Injection for Application-level Reliability Improvement of Machine Learning-based Computing

摘要

In conventional SSD controller, error correcting code (ECC) is required to correct errors in 3D-TLC NAND flash memories. However, in machine learning (ML)-based computing, some errors can be tolerated and ECC is not necessary. By utilizing the error tolerance, this paper proposes Approximate Initial Error Injection (AIEI) which initially injects errors to memory cells to improve the application-level reliability after the data-retention (D.R.). AIEI initially injects 16.7% of errors to 3D-TLCNAND flash cells. Then, after the D.R., measured bit-error rate (BER) decreases by 34%. Measured acceptable D.R. time is extended by 4.9-times.