12 dec. 2008

How We Found the Missing Memristor By R. Stanley Williams

The fourth elementary unit to construct a programmable brain-like computer is the memristor. The memristor — the functional equivalent of a synapse — could revolutionize circuit design. First: read the article.

Emulating the behavior of a single memristor, Chua showed, requires a circuit with at least 15 transistors and other passive elements. The implications are extraordinary: just imagine how many kinds of circuits could be supercharged by replacing a handful of transistors with one single memristor.

The most obvious benefit is to memories. In its initial state, a crossbar memory has only open switches, and no information is stored. But once you start closing switches, you can store vast amounts of information compactly and efficiently. Because memristors remember their state, they can store data indefinitely, using energy only when you toggle or read the state of a switch, unlike the capacitors in conventional DRAM, which will lose their stored charge if the power to the chip is turned off. Furthermore, the wires and switches can be made very small: we should eventually get down to a width of around 4 nm, and then multiple crossbars could be stacked on top of each other to create a ridiculously high density of stored bits.

Greg Snider and I published a paper last year showing that memristors could vastly improve one type of processing circuit, called a field-programmable gate array, or FPGA. By replacing several specific transistors with a crossbar of memristors, we showed that the circuit could be shrunk by nearly a factor of 10 in area and improved in terms of its speed relative to power-consumption performance. Right now, we are testing a prototype of this circuit in our lab.

And memristors are by no means hard to fabricate. The titanium dioxide structure can be made in any semiconductor fab currently in existence. (In fact, our hybrid circuit was built in an HP fab used for making inkjet cartridges.) The primary limitation to manufacturing hybrid chips with memristors is that today only a small number of people on Earth have any idea of how to design circuits containing memristors. I must emphasize here that memristors will never eliminate the need for transistors: passive devices and circuits require active devices like transistors to supply energy.

The potential of the memristor goes far beyond juicing a few FPGAs. I have referred several times to the similarity of memristor behavior to that of synapses. Right now, Greg is designing new circuits that mimic aspects of the brain. The neurons are implemented with transistors, the axons are the nanowires in the crossbar, and the synapses are the memristors at the cross points. A circuit like this could perform real-time data analysis for multiple sensors. Think about it: an intelligent physical infrastructure that could provide structural assessment monitoring for bridges. How much money—and how many lives—could be saved?

I’m convinced that eventually the memristor will change circuit design in the 21st century as radically as the transistor changed it in the 20th. Don’t forget that the transistor was lounging around as a mainly academic curiosity for a decade until 1956, when a killer app—the hearing aid—brought it into the marketplace. My guess is that the real killer app for memristors will be invented by a curious student who is now just deciding what EE courses to take next year.