Nanosilver and CNT-nanocomposite toxicology in an in vivo model, D. Melanogaster

Kyle Robert Murphy

Abstract

The memristor is a novel nano-scale device discovered in 2008. Memristors are basically nonvolatile variable resistors. Various breakthroughs of memristive devices have shown the potential of memristive crossbar designs for their ultra-high density and low-power memory. Initial studies have shown that memristor based neuromorphic processors could potentially consume 300,000 times less power than a traditional Intel Xeon Processor for neural network applications. These neuromorphic processors require large memristor arrays for evaluating neural networks. A key problem in designing these processors is simulating the large arrays of thousands of memristors as these simulations require extremely long times (sometimes weeks to months). Additionally, most existing simulation tools are unable to handle the large number of computations seen in these simulations. The simulations of these large arrays has not been examined in literature. One of the key objectives of this work is to evaluate large scale memristor crossbars that allow high density layout of synapses and thus enable building of highly capable neuromorphic systems. In order to achieve this, this study utilized a newly released parallel SPICE simulator, Xyce, developed by Sandia National Labs. Although it is 4 times faster than the existing SPICE simulators, Xyce is still too slow for evaluating the capabilities of large crossbar arrays. Therefore, this thesis also examines the development of an equivalent mathematical representation of the circuit level memristor based neuromorphic circuits for faster computation of the crossbars without the use of SPICE. This system is then used as an offline training approach to approximate the SPICE circuit. These trainings require thousands of iterations of the crossbar array. This thesis examined the training of memristor-based crossbars for neural network based and pattern recognition applications such as Image classification at the circuit level. Finally the analysis presented in this work will be crucial in understanding the future of the memristor-based crossbars in developing highly reliable and extremely low power processors and neuromorphic systems.