UC San Diego

As conventional hard decision error correction codes (ECCs), such as BCH codes, become inadequate as the capacity of flash memory increases, soft decision ECCs such as LDPC codes have become a desirable option. However, fully utilizing the potential of soft decision based codes demands higher precision memory sensing with the tradeoff of memory read latency. To that extent, this thesis explores and compares two approaches to optimizing the positioning as well as the number of read (word-line) voltages for a specified program/erase (PE) cycle.

The first approach is optimization subject to maximizing the mutual information (MMI) of the equivalent discrete memoryless channel using gradient descent. We show that we can get a near optimal performance by only using ~20 reads since additional reads beyond this point provides less than 0.05% of additional mutual information and less than 5% of additional performance gain.

The second approach is the optimization of read thresholds taking the code structure of LDPC codes into account. We use discretized density evolution as a proxy for bit error rate, and to serve as our performance measure in the gradient descent search. This optimization is code-dependent and gives lower BER; however, it is subject to be trapped in secondary local minimum. As a result, we propose a two-step optimization: first utilize the MMI approach for coarse read threshold optimization, then follow by the DE-based method for fine optimization.