Jan 26, 2012
Analog memory paves the way for efficient information processing
The past several decades have been marked by the exponential growth of computer-generated data and related information processing. Such growth continues, for example with the deployment of gigabit internet and 4G wireless networks, and will likely be accelerated by emerging technologies such as robotics, biotechnology and distributed sensor networks. Given the inevitable end of the scaling of conventional semiconductor (CMOS) circuits and increasing awareness of energy use, alternative ways to allow for information processing in an energy-efficient fashion must be developed. Recently, Fabien Alibart and colleagues at the University of California Santa Barbara, US, have demonstrated analog memory that presents a key milestone towards information processing with hybrid circuits, which have the potential to introduce a paradigm shift over current practice.
In many cases, and especially for emerging applications, information processing at relatively low (less than 8-bit) precision could be sufficient. The most frequent operation, and as a consequence the bottleneck, in this context is typically a dot product or Multiply-and-Add Computation (MAC). For example, image processing from a focal plane array of sensors relies heavily on spatial and temporal filtering and edge detection, which involve the computation of convolutions or correlations, that is MACs, on 8-bit pixels. Likewise, information from sensors attached to the human body, for example to monitor the wearer's heart beat, or from sensors deployed on the sea bed for timely prediction of earthquakes, may be limited to low precision without affecting the performance of the circuit, and would typically involve MAC-based compressed sensing, filtering and/or bio-inspired neuromorphic information processing.
Theoretically (that is from physical noise analysis) any computation at such low precision could be done much more efficiently in analog (or mixed signal) circuits, with wires carrying multiple bits of information. However, even low-precision analog processing in conventional CMOS technology is cumbersome due to the lack of suitable hardware.
The team led by Prof. Strukov is looking at ways to perform low-precision analog information processing with hybrid circuits that combine single conventional CMOS chip and layer(s) of quasi-passive resistive switching ("memristor") nanoscale crosspoint devices. The main advantages of memristors are their very high integration density, and the fact that they can be switched between high and low resistive states continuously, reversibly and in a nonvolatile fashion.
To realize MAC circuits, the memristive devices implement density-critical configurable low-precision weights, while CMOS is used for the summing amplifier, which provides gain and signal restoration. As a result, individual voltages applied to memristors can be multiplied by the unique weight (conductance) of the memristor and summed by the CMOS amplifier – all in analog fashion.
Using specific dynamic properties of the memristive devices, Strukov's group has developed a technique to achieve 7-bit precision operations, even in the presence of intrinsic variations in the memristive devices and, in particular, has demonstrated two-input 7-bit hybrid MAC circuitry. These results are very encouraging and present a major milestone towards implementation of hybrid circuits.
Additional information can be found in the journal Nanotechnology.
About the author
The research was conducted by Dr F Alibart, Dr L Gao, B Hoskins and Prof. D Strukov from UC Santa Barbara and is supported by the US National Science Foundation.