Advances in Computing Technology: Memristors and Neural Applications
October 30th, 2013
Rapid advances in computing technology have depended upon Moore's Law, which refers to the ability to scale down transistor size with each generation of technology. Due to physical limitations, this scaling is now in jeopardy for memory technologies. Memristor is a new memory technology based upon electrical resistance, in contrast to traditional Dynamic Random Access Memory (DRAM) technology, which is based upon electrical charge. Memristors avoid many of the scaling challenges of DRAM and are expected to be utilized as a potential scalable alternative for DRAM. Memristor cells can be used to build high density memory that can retain its information even when it loses power. In addition, these memristor cells can be used to perform computations as well.
Memristors for neural applications
Memristors and neural applications have always been considered a perfect match. The strong affinity between them is generally based upon the assumption that memristors would be used to mimic a complex network of neurons. Unfortunately, this purely analog implementation is impractical given the serious issues that currently remain with the technology, such as low write endurance and high defect rates. We find that a hybrid digital-analog approach is needed. In addition, we have exploited memristors’ unique analog properties to develop general techniques that can leverage devices with significantly low durability (for example, devices that can survive only a few hours of continuous switching). These new techniques can provide a system lasting for five or more years of continuous operation.
The intended long-term use of this technology is to simulate the brain, as they are expected to provide the form factor of a brain, the low power requirements, and the instantaneous internal communications. Our research referenced below is an initial step towards integrating the memristor technology into challenging environments starting from wireless sensor networks and up to highly updated neural nets such as the brain.
What are the applications of this research?
An important issue to explore in future studies is to examine closely various neural applications such as face recognition, classification, and misuse detection (detecting attacks on vulnerable data). These applications rely on a cross-product circuit that can efficiently be implemented using hybrid Memristor-CMOS circuits. We are now working on identifying the potential benefits of high density for such applications, as well as evaluating the technology limitations, and the energy savings of using memristors for such applications.
The main purpose of our research is to enable energy-efficient implementation of neural-network applications using hybrid Memristors-CMOS circuits. Furthermore, we develop techniques that enable these memories to be adopted in challenging environments.
Author: Hebatallah Saadeldeen, October 2013
Department of Computer Science, UC Santa Barbara
To read full paper on research findings click here.
Authors and publication
Hebatallah Saadeldeen, Diana Franklin, Guoping Long, Charlotte Hill, Aisha Browne, Dmitri Strukov, Timothy Sherwood, and Frederic T. Chong, "Memristors for Neural Branch Prediction: A Case Study in Strict Latency and Write Endurance Challenges." ACM International Conference on Computing Frontiers Article No. 26 (2013).