Scientists Create First Memristor: Missing Fourth Electronic Circuit Element
By Bryan Gardiner EmailApril 30, 2008 | 12:03:41 PMCategories: Research
Memristor
Researchers at HP Labs have built the first working prototypes of an important new electronic component that may lead to instant-on PCs as well as analog computers that process information the way the human brain does.
The new component is called a memristor, or memory resistor. Up until today, the circuit element had only been described in a series of mathematical equations written by Leon Chua, who in 1971 was an engineering student studying non-linear circuits. Chua knew the circuit element should exist -- he even accurately outlined its properties and how it would work. Unfortunately, neither he nor the rest of the engineering community could come up with a physical manifestation that matched his mathematical expression.
Thirty-seven years later, a group of scientists from HP Labs has finally built real working memristors, thus adding a fourth basic circuit element to electrical circuit theory, one that will join the three better-known ones: the capacitor, resistor and the inductor.
Researchers believe the discovery will pave the way for instant-on PCs, more energy-efficient computers, and new analog computers that can process and associate information in a manner similar to that of the human brain.
According to R. Stanley Williams, one of four researchers at HP Labs' Information and Quantum Systems Lab who made the discovery, the most interesting characteristic of a memristor device is that it remembers the amount of charge that flows through it.
Indeed, Chua's original idea was that the resistance of a memristor would depend upon how much charge has gone through the device. In other words, you can flow the charge in one direction and the resistance will increase. If you push the charge in the opposite direction it will decrease. Put simply, the resistance of the devices at any point in time is a function of history of the device –- or how much charge went through it either forwards or backwards. That simple idea, now that it has been proven, will have profound effect on computing and computer science.
"Part of what's going to come out of this is something none of us can imagine yet," says Williams. "But what we can imagine in and of itself is actually pretty cool."
For one thing, Williams says these memristors can be used as either digital switches or to build a new breed of analog devices.
For the former, Williams says scientists can now think about fabricating a new type of non-volatile random access memory (RAM) – or memory chips that don't forget what power state they were in when a computer is shut off.
That's the big problem with DRAM today, he says. "When you turn the power off on your PC, the DRAM forgets what was there. So the next time you turn the power on you've got to sit there and wait while all of this stuff that you need to run your computer is loaded into the DRAM from the hard disk."
With non-volatile RAM, that process would be instantaneous and your PC would be in the same state as when you turned it off.
Scientists also envision building other types of circuits in which the memristor would be used as an analog device.
Indeed, Leon himself noted the similarity between his own predictions of the properties for a memristor and what was then known about synapses in the brain. One of his suggestions was that you could perhaps do some type of neuronal computing using memristors. HP Labs thinks that's actually a very good idea.
"Building an analog computer in which you don't use 1s and 0s and instead use essentially all shades of gray in between is one of the things we're already working on," says Williams. These computers could do the types of things that digital computers aren't very good at –- like making decisions, determining that one thing is larger than another, or even learning.
While a lot of researchers are currently trying to write a computer code that simulates brain function on a standard machine, they have to use huge machines with enormous processing power to simulate only tiny portions of the brain.
Williams and his team say they can now take a different approach: "Instead of writing a computer program to simulate a brain or simulate some brain function, we're actually looking to build some hardware based upon memristors that emulates brain-like functions," says Williams.
Such hardware could be used to improve things like facial recognition technology, and enable an appliance to essentially learn from experience, he says. In principle, this should also be thousands or millions of times more efficient than running a program on a digital computer.
The results of HP Labs teams findings will be published in a paper in today's edition of Nature. As far as when we might see memristors actually being used in actual commercial devices, Williams says the limitations are more business oriented than technological.
Ultimately, the problem is going to be related to the time and effort involved in designing a memristor circuit, he says. "The money invested in circuit design is actually much larger than building fabs. In fact, you can use any fab to make these things right now, but somebody also has to design the circuits and there's currently no memristor model. The key is going to be getting the necessary tools out into the community and finding a niche application for memristors. How long this will take is more of a business decision than a technological one."
Exquisite! Starglider's scheme appears to be moving along nicely.
Great, now I have to learn about another circuit element...
Just kidding, this kicks ass and will be great for future applications in pretty much everything, because building better computers means building better everything.
I love learning. Teach me. I will listen. You know, if Christian dogma included a ten-foot tall Jesus walking around in battle armor and smashing retarded cultists with a gaint mace, I might just convert - Noble Ire on Jesus smashing Scientologists
The component design is great, though it sounds like it's a fair way from mass production. I'm underwhelmed by the notion of using it to implement ANNs directly; this idea has been tried before with less exotic components and it didn't work terribly well. The basic problem is that our existing training rules (operating on very simple neuron models) don't scale and no one really knows how the brain (which has much more sophisticated neuron behaviour than a typical ANN) does it yet. I imagine this could make simple NN based apps run on less power, but that would be an incremental improvement on a FPGA implementation. It also risks suffering from the same problem as bubble memory and holographic storage; new technology that's theoretically better but never catches up with 'conventional' technology in practice due to much more money being thrown at refining the manufacturing tech for the later. For notions of emulating the whole brain; until we have a good model running in software, any attempt to implement in hardware is basically premature optimisation. The whole Hugo de Garis brain machine fiasco demonstrated this quite amusingly. As for analogue computing, we ditched it back in the 50s despite very tight component counts for several very good reasons; it's a whole load of hassle at every level and the benefits generally aren't worth the costs.
