Memristive Hardware Artificial Neural Networks – MHANN

In 2008, researchers at Hewlett-Packard have unveiled a new electronic component, called the memristor. Theorized by Leon Chua in 1971, this electronic device is a non-volatile, non-linear resistor. By applying a voltage, it is possible to vary continuously the resistance of the device, and the device "memorizes" that resistance after the voltage is no longer applied. As such, these memristors offer a wide variety of applications as binary (OFF/ON) or multilevel/analog memories or switches in reconfigurable memories. In addition, memristors intrinsically behave as artificial synapses. Most of the memristor devices proposed up to this day are based on defect-mediated physical effects, for example the electromigration of oxygen ions for the Hewlett-Packard components. To use such devices, the reliability issues due to high operating temperatures, difficulty of a precise control of the switching behavior and potential device deterioration will have to be solved in order to achieve the large endurance and retention times required for operational components.
In 2009, the UM?-CNRS and Thales project partners have patented a new component: the “ferroelectric memristors”. This memristor belongs to another class of memristors, called “Electronic effect memories”. The resistance changes are due to purely electronic effects and therefore preserve the materials structure. They are based on a physical concept radically different from the existing solutions: ferroelectricity in tunnel junctions. The resistive switching is based on the intrinsic switching of ferroelectric domains and therefore possesses a fundamental merit over defect-mediated mechanisms to achieve the reliable performance necessary for commercial production.
On the other hand, companies like Intel have insisted that the most important high-performance applications are not scientific computing but the following three categories: Recognition, Mining and Synthesis applications (RMS), the first two categories largely relying on classification, clustering, approximation and optimization algorithms, for which competitive algorithms based on neural networks exist. Due to stringent power consumption constraints, the clock frequency of processors no longer increases or barely, so that a hardware neural network would retain all its performance and power advantage (about two orders of magnitude) compared to the software version run on a processor.
So we can note that there is a convergence of technology, architecture and application trends: hardware artificial neural networks are well suited to tackle an important class of applications while coping with upcoming technology hassles.
Therefore, memristors constitute an ideal and very timely alternative implementation for synapses of hardware artificial neural networks. Memristors would be far denser than current SRAM-based implementations but they would also require far less power since the memristor is a non-volatile memory. Hardware ANNs with an architecture composed of analog circuitry coupled with the aforementioned memristors open the possibility to build high-performance accelerators able to tackle the large computational tasks of RMS applications.
The purpose of this project is to build a medium-scale prototype of such a bio-inspired architecture, by using long-life and nanometric “ferroelectric” memristors. The area, performance and power benefits of this approach will be evaluated to define its interest for embedded systems.
The MHANN project is multi-disciplinary in the sense that it proposes new physical concepts for devices (physics) and aims at integrating them into on-chip bio-inspired architectures (micro-electronics, computer science and architectures).