|M.Sc Student||Ben-Hur Yuval|
|Subject||Detection and Coding Schemes for Resistive Storage Devices|
|Department||Department of Electrical Engineering||Supervisor||Professor Yuval Cassuto|
|Full Thesis text|
The amount of digitally formatted data created by mankind has grown exponentially over the last few decades. This trend line fuels constant efforts to build denser and more reliable storage devices. The ongoing demand for better performance requires data systems to constantly accelerate their operation, consume less power and miniaturize their dimensions. Even with the sophisticated technologies developed so far, all of these requirements are still relevant today, as leading technologies, such as Flash and DRAM, seem to reach a scaling barrier.
One of the most exciting technologies to emerge recently was the Memristor. Memristors are a general term for several resistive-based storage and computing cell technologies. The key in all those technologies is that the memory cell is a passive two-terminal device that can be both read and written over a simple crossbar structure. This feature offers a huge density advantage, but at the cost of poor isolation between cells, resulting in severe access and reliability issues. Mitigating these issues is a highly motivated objective, given the far-reaching impact resistive arrays can strike on future computing systems.
This Master thesis focuses on a fundamental issue associated with the crossbar architecture for resistive memories. During a cell readout, alternative current tracks consisting of array cells other than the target cell are formed. These tracks, called sneak paths, highly influence the measurement of a target cell. Up to this date, sneak paths were addressed mostly as a hardware problem. This approach lead to solutions in the architecture level, which come on the expense of either memory density or read latency.
We study the problem from a communication-theory perspective, conceptualizing sneak paths as a noisy channel with inter-cell interference. The channel is modeled in a way that captures the unique physical effects caused by sneak paths. Optimal Maximum A-Posteriori detectors for cell readout over the sneak path channel are developed and evaluated. They are shown to dramatically improve data reliability compared to standard readout techniques. An additional improvement in performance is achieved by using multiple measurements of the same cell. We also develop a simplified version of the optimal detector, which consists of a two-step estimation scheme that is substantially easier to implement. This sub-optimal detector is shown to achieve results comparable to the optimal schemes in terms of bit error rate.
Next, we develop a low-complexity detector that simply compares the measurement to a predefined threshold. We estimate the optimal threshold based on approximations that stem from a deep understanding of the unique setup of the sneak path model. The simpler threshold detector is shown to perform close to the optimal threshold in bit error rate while significantly reducing the computational complexity of the detection process, also for multiple measurements. Finally, a further enhancement of readout reliability is achieved by a data-encoding technique that suppresses the incidence of sneak paths within the array. The combination of mapping the data and jointly detecting the entire code-word leads to even better performance.