Dynamic Error Recovery Flow Prediction Based on Reusable Machine Learning for Low Latency NAND Flash Memory Under Process Variation

Abstract

NAND flash memory is becoming smaller and denser to have a larger storage capacity as technologies related to fine processes are developed. As a side effect of high-density integration, the memory can be vulnerable to circuit-level noise such as random telegraph noise, decreasing the reliability of the memory. Therefore, low-density parity-check code that provides multiple decoding modes is adopted in the NAND flash memory systems to have a strong error correcting capability. Conventional static error recovery flow (ERF) applies decoding modes sequentially, and read latency can increase when preceding decoding modes fail. In this paper, we consider a dynamic ERF using machine learning (ML) that predicts an optimal decoding mode guaranteeing successful decoding and minimum read latency and applies it directly to reduce read latency. Due to process variation incurred in the manufacturing of memory, memory characteristics are different by chips and it becomes difficult to apply a trained prediction model to different chips. Training the customized prediction model at each memory chip is impractical because the computational burden of training is heavy, and a large number of training data is required. Therefore, we consider ERF prediction based on reusable ML to deal with varying input and output relationships by chips due to process variation. Reusable ML methods reuse pre-trained model architecture or knowledge learned from source tasks to adapt the model to perform its task without any loss of performance in different chips. We adopt two reusable ML approaches for ERF prediction based on transfer learning and meta learning. Transfer learning method reuses the pre-trained model by reducing domain shift between a source chip and a target chip using a domain adaptation algorithm. On the other hand, meta learning method learns shared features from multiple source chips during the meta training procedure. Next, the meta-trained model reuses previously learned knowledge to fastly adapt to the different chips. Numerical results validate the advantages of the proposed methods with high prediction accuracy in multiple chips. In addition, the proposed ERF prediction based on transfer and meta learning can yield a noticeable reduction in average read latency as compared to conventional schemes.INDEX TERMS NAND flash memory system, process variation, error management, reusable machine learning, transfer learning, meta learning.

Read Reference Voltage Adaptation for NAND Flash Memories With Neural Networks Based on Sparse Histograms

“…These define three voltage intervals. In [28], a similar figure is generated, called the oncell ratio, where the number of cells turned on at a reference voltage is divided by the total number of cells.…”

“…In order to reduce the average read and decoding latency [28] presented an approach to predict the optimal decoding method called dynamic error recovery flow. This method does not predict optimal RRVs, rather it assumes optimal RRVs.…”

“…The on-cell ratio is the ratio between active and total cells at a fix predefined RRV. Note that for the classification in [28], two on-cell ratios at different RRV are required.…”

“…The features of the NN are sparse histograms. The sparse histogram component is similar to the description of the on-cell ratio from [28]. However, we do not calculate a ratio, instead we count the active cells between adjacent RRVs, and it is not strictly necessary to use predefined RRVs.…”