Abstract
In recent years, non-volatile memories (NVMs) such as the flash memory have found a wide range of applications, ranging from consumer electronics to large scale data centres. However, physical and technological constraints make the flash memory hard for further scaling. Emerging NVM technologies such as the spin-torque trans fer magnetic random access memory (STT-MRAM) and resistive RAM (ReRAM) have shown great potential to be the next generation NVMs. Irrespective of their attrac tive features, undesirable noise and interference severely degrade the reliability of data stored in these emerging memories. This thesis is devoted to the design and analysis of novel channel coding and detection schemes to tackle the various channel impairments and improve the reliability of emerging NVMs. First, to mitigate the write errors and read errors caused by the thermal fluctuation and process variation of STT-MRAM, this thesis proposes novel rate-compatible pro tograph low-density parity-check (RCP-LDPC) codes. The proposed codes with rate compatible property can reduce the raw bit error rate (BER) diversity of the memory and outperform the state-of-the-art fixed-rate Bose-Chaudhuri-Hoquenghem (BCH) codes and Euclidean Geometry (EG) LDPC codes. As high-precision analog-to-digital converters (ADCs) are not applicable for high speed STT-MRAM, the design of channel quantizer is critical to support the error cor rection codes (ECCs) for STT-MRAM. This thesis first proposes a union bound analysis which can accurately predict the word error rates (WERs) of ECCs with maximum like- v lihood decoding over the quantized STT-MRAM channel, without resorting to lengthy computer simulations. Moreover, leveraging on the union bound analysis and differen tial evolution algorithm, this thesis further presents a union-bound-optimized quanti zation scheme which outperforms the prior-art quantization schemes for STT-MRAM. The memory physics induced unknown offset is yet another critical and difficult issue for STT-MRAM. This thesis investigates and presents a neural network (NN)- based dynamic threshold detection (DTD) scheme to tackle it. The main idea is to adjust the threshold of the detector dynamically based on the outputs of the NN detec tor developed for STT-MRAM. The proposed detection scheme can achieve the perfor mance of the optimum detector without the prior knowledge of the channel. Moreover, a novel deep learning (DL)-based neural normalized-offset reliability-based min-sum (NNORB-MS) decoding algorithm is further proposed to improve the decoding perfor mance of linear block codes for STT-MRAM. In the ReRAM crossbar array, the sneak path interference (SPI) will lead to strong inter-cell correlation which affects the data reliability severely. Due to the complication of memory physics and unique features of the SPI, it is difficult to derive an accurate channel model for it. This thesis proposes a novel constrained coding (CC) and DL aided threshold detection scheme to combat the SPI, without the prior knowledge of the channel. The proposed CC scheme can not only reduce the SPI in the memory array, but also effectively differentiate the memory arrays into two categories. This enables the application of two different threshold detectors for memory arrays without and with SPI, respectively. The threshold detector for the SPI-affected arrays are designed using the outputs of a NN detector developed for ReRAM. Lastly, the existence of the SPI and the read disturb induced unknown channel off set makes the ReRAM channel unstable, and its channel raw BER (i.e. the channel condition) changes with different input data and/or different number of reads. This thesis proposes a novel DL-based adaptive decoding algorithm of ECCs, whose de coding complexity can be adjusted according to the change of the channel condition of ReRAM with a single encoder and decoder, so as to save the energy consumption and vi reduce the read latency.