A Comparative Study of Classification Methods for Flash Memory Error Rate Prediction

Abstract

NAND Flash memory has been the fastest growing technology in the history of semiconductors and is now almost ubiquitous in the world of data storage. However, NAND devices are not error-free and the raw bit error rate (RBER) increases as devices are programmed and erase (P-E cycled). This requires the use of error correction codes (ECCs), which operate on chunks of data called codewords. NAND manufacturers specify the number of P-E cycles a device can tolerate (known as endurance) very conservatively to account for quality variations within and across devices. This research uses machine learning to predict the true cycling level each part of a NAND device can tolerate, based on measurements taken from the device as it is used. Real data is gathered on millions of codewords and eight machine learning classification methods are compared. A new subsampling method based on the error probability density function is also proposed.