Deep Geek: Understanding Memristors

December 7, 2008 7:54 PM

The coming memristor revolution in electronics and how it works.The newly created memristor, only the fourth fundamental fundamental type of passive circuit element, has the promise of computing advances both prosaic (faster, cheaper and "bigger" flash drives) and momentous (relatively effortless mimicry of brain cells and their activity). This is the story of the memristor's genesis, told by R. Stanley Williams, the leader of the team that created the device.

Being deeply geeky myself, I've read about memristors before, but reading this article and sidebar finally let me understand how the memristor works and what happens inside it. And that felt pretty damn good.

The article is fantastic, but it does leave one key connection unmade. To create a practical memristor, the team "needed [a] mechanism by which we could change the effective spacing between two wires in our crossbar by 0.3 nm. If we could do that, we would have the 1000:1 [variation in conductivity] we needed... Where would we find a material that could change its physical dimensions like that?" They did create a way to vary that spacing, in a controllable, repeatable, and extremely fast-acting manner, but Williams doesn't directly explain how the internal actions of the switching layer meet that requirement. The payoff for that setup is missing.

When electrical current pushes the conductive impurities in the layer of titanium dioxide toward the other wire, the conductive portion of the layer grows toward the other wire, and the insulating portion of the layer thins.

That thinning is described, but the article never tells the reader that expansion of the conductive layer is that long-sought means of moving the wires.

If you read the article and made that connection before I described it, then you might have felt as smug about it as I did. Williams gets to feel more smug.

posted by NortonDC (43 comments total)

Very cool. I just wish I could understand even the first page of this article. Is this an example of the much hyped quantum computing?
posted by zardoz at 8:08 PM on December 7, 2008

Memristor at wikipedia.
posted by Pastabagel at 8:15 PM on December 7, 2008

Wow, really interesting article NortonDC.

One of the other interesting things he talks about is his effort at HP to build a supercomputer composed of defective parts that would keep working. This reminds me of what I read about the Soviet chip fabrication industry, which was evidently dead by the early eighties because of insufficient quality control - they had to ship a list of the instructions that didn't work along with each chip and software had to be re-written to account for the hardware defects.

It would be an awesome basis for an alternative history novel, if some Soviet scientist in the seventies figured out a similar approach to compensate for a high defect rate...
posted by XMLicious at 8:24 PM on December 7, 2008

This is not quantum computing. This is not as massive a leap as quantum computing would be, except that this is actually real, and buildable now with current fabrication technology.

The best way to understand the importance of this is to consider that there are four fundamental electrical properties at work in a circuit: voltage, current, charge and flux. The known fundamental circuit elements, resistors, inductors, and capacitors, manipulate those forces in some way. For example, a resistor has a resistance which is simply voltage divided by current. By contrast, a transistor is not one of these, because it is basically doing what a network of simpler circuit elements is doing. In other words, a transistor can be modeled by a circuit of only resistors, capacitors, and inductors.

The memristor relates two of those that were previously unrelatable, charge and flux. There is no resistor, inductor, capacitor equivalent circuit for a memristor.

Coupling the memristor with the crossbar latch structure is going to revolutionize solid state memory. This is a very big deal, and unlike quantum computing, this might actually show up in the real world in 5-10 years.
posted by Pastabagel at 8:35 PM on December 7, 2008

By contrast, a transistor is not one of these, because it is basically doing what a network of simpler circuit elements is doing. In other words, a transistor can be modeled by a circuit of only resistors, capacitors, and inductors.

Are you certain of this? It's been a long time since I've studied any of this stuff, but I was damn near certain that nonlinear devices (e.g. transistors and diodes) could not be modeled by passive networks (of resistors, capacitors, and inductors). At least, I should say, that a given network including nonlinear devices could not be represented with the same passive network for all possible operating points (of biases and frequencies).
posted by Jpfed at 8:49 PM on December 7, 2008

So. fucking. cool.
posted by Pope Guilty at 8:54 PM on December 7, 2008

Pastabagel: In what way could a (finite) network of passives be like a transistor?
posted by phrontist at 8:56 PM on December 7, 2008

So this works because voltage causes "bubbles of oxygen deficiency" to migrate into and out of the pure TiO2 layer, causing persistent, non-volatile changes in the switch's conductivity. How can this be made to happen? At GHz speed?
posted by George_Spiggott at 9:04 PM on December 7, 2008

Argh. That question should read "How quickly can this be made to happen?"
posted by George_Spiggott at 9:05 PM on December 7, 2008

In other words, a transistor can be modeled by a circuit of only resistors, capacitors, and inductors.

A nonlinear device such as a transistor cannot be reproduced by a combination of linear passive devices.

Its not immediately clear to me why these are so potentially useful. Capacitors can store voltage in just the same way that memristors seem to be able to store resistance, so the implication that other passive components don't have 'memory' is false. I suspect that the impact is more of a technological improvement than a fundamental leap. Meaning these might be faster, cheaper, more efficient - but I doubt they are able to do things that we haven't already been able to do with existing components.
posted by jpdoane at 9:24 PM on December 7, 2008

Pastabagel: you are incorrect. Actives and passives are different. Inductors, capacitors, and resistors can perform only linear operations. Transistors and diodes and the like are active components. They are the source of (practically) all nonlinearity, which is of course the stuff of interest in electronics.

I also don't know where you get these ideas about charge and [magnetic?] flux being "unrelatable," whatever that means. You can build a circuit to connect any two arbitrary quantities according to any arbitrary formula. A memristor can certainly be modeled by an equivalent circuit. For example: we can simulate them on computers, and a computer is a (big) transistor circuit. Throw an ADC and a DAC in there and you have an equivalent circuit (and yes, not that it matters, but you could also do it all analog if you wanted).
posted by Xezlec at 9:29 PM on December 7, 2008

Because memristors behave functionally like synapses, replacing a few transistors in a circuit with memristors could lead to analog circuits that can think like a human brain.

Barf.
posted by zippy at 9:39 PM on December 7, 2008 [6 favorites has favorites]

Very interesting. One slightly worrying issue, although I'm sure it'll be resolved by the time these are ready for commercial production: sometimes one wants to switch a computer off and on, especially (but not exclusively) if it is running Windows, to clear its memory, to kill off crashed programs that have hung the computer.

So it would be useful to have at least three off-and-on states, I think: factory default, at least one user-editable "safe state" analogous to the current bootup state, and "awake from sleep" mode. These do work under the current model, but not as well as this technology would.
posted by aeschenkarnos at 9:41 PM on December 7, 2008

OK, I see where the charge/flux thing came from, it was Wikipedia. They meant electric flux, and they meant that those two quantities within the device itself were physically connected.
posted by Xezlec at 9:43 PM on December 7, 2008

Oh, no they did mean magnetic flux. I'll stop now before I screw anything else up.
posted by Xezlec at 9:46 PM on December 7, 2008

Its not immediately clear to me why these are so potentially useful.

The point, as far as I can tell, is that memristors represent a physically-compact and high-performance way of enforcing electrical constraints of interest.

The English language could get by without the use of certain nouns; every idea currently expressible could still be expressed. But using those nouns lets us express ourselves in less space (or time). For all the time we've been developing electronics, we've been using a somewhat impoverished vocabulary, and so taken more time and space to achieve the desired effects.

Capacitors can store voltage in just the same way that memristors seem to be able to store resistance, so the implication that other passive components don't have 'memory' is false.

Well, yes, but very small capacitors suck. If I remember correctly, capacitance is proportional to the area of the surfaces in question and inversely proportional to their distance. At nano-scales, this is not favorable (distance becomes small, but area becomes VERY small). The articles implied that memristance(?) scaled favorably.
posted by Jpfed at 9:49 PM on December 7, 2008

Because memristors behave functionally like synapses, replacing a few transistors in a circuit with memristors could lead to analog circuits that can think like a human brain.

If they want it to think like my brain, they're going to have to find a way to add beeristors, too..
posted by stavrosthewonderchicken at 9:53 PM on December 7, 2008

In other words, a transistor can be modeled by a circuit of only resistors, capacitors, and inductors. -- pastabagel.

Wow, that's so wrong, as other people have already pointed out. But I just wanted to repeat it because it was so shockingly wrong.

Anyway, about the article. Wow, this is setting off a lot of "bad science writing" alarm bells for me. Particularly the "revolutionary" aspect. What's so revolutionary about this? We already have single components that can store data, such as 1t-SRAM, a type of memory that uses just one transistor to store one bit of data. There's also Z-RAM which is cool because rather then a capacitor, it uses the properties of newer SOI (Silicon on Insulator) fabrication to get a sort of "free" capacitor.

What they describe isn't exactly like a synapse, either. A synapse isn't even a brain cell, a synapse is the gap between braincells. Synapses don't behave anything like this memristors. They do have a sort of time-dependant memory effect, when a synapse fires it loses some of it's stored neurotransmitter, so if you fire it too much in too short of an amount of time, it won't be able to fire until a 'refractory period' is complete.

So, I guess you could have one that would shut off after a certain amount of electricity was sent through, but you would have to have it open up after a certain amount of time, to simulate the replenishment of the neurotransmitter.

But couldn't you do the same thing with a capacitor?
posted by delmoi at 9:56 PM on December 7, 2008

How can this be made to happen? At GHz speed?

Heck no. Wikipedia says it's only been done at 1 Hz.

I doubt this is all that important, really. It's just yet another new memory technology that's probably slower than flash memory. The only reason it's interesting is that it completes the pretty little mathematical picture of four passive circuit elements.

Well, yes, but very small capacitors suck. If I remember correctly, capacitance is proportional to the area of the surfaces in question and inversely proportional to their distance. At nano-scales, this is not favorable (distance becomes small, but area becomes VERY small).

Trench caps (i.e. doing microprocessors with a DRAM process) are probably easier than memristors. That gives you pretty substantial area per cap. Anyway, I don't quite agree with you about capacitor scaling. Although capacitance scales down linearly, transistor sensitivity also scales up accordingly. So it basically stays the same. Small capacitors don't suck. Memristors suck. I mean they physically move ions around. Come on, that'll never be fast.
posted by Xezlec at 10:00 PM on December 7, 2008

...so if you fire [a synapse] too much in too short of an amount of time, it won't be able to fire until a 'refractory period' is complete.

Whoa. So we need brain viagra?
posted by rokusan at 10:01 PM on December 7, 2008

Although capacitance scales down linearly, transistor sensitivity also scales up accordingly.

Touche.
posted by Jpfed at 10:20 PM on December 7, 2008

Whoa. So we need brain viagra?

Well, we have plenty of drugs that dink around with synapses. Selective serotonin reuptake inhibitors, for example work by keeping the neurotransmitter (in this case, serotonin) in the synapse longer, which can cause the recipient neuron to receive it over and over again. Ecstasy works in a similar way.
posted by delmoi at 10:39 PM on December 7, 2008

Excuse me if this is a dumb question, but how do you read the stored state of the thing? You set it by applying a certain current at a certain polarity. It then 'preserves' a record of that data. But how do you read it without changing it again?
posted by woodblock100 at 11:26 PM on December 7, 2008 [1 favorite has favorites]

Xezlec: Memristors suck. I mean they physically move ions around. Come on, that'll never be fast.

Is that really true or have I misinterpreted this? I would defer to your knowledge of it, since you clearly have more experience at this stuff than me, but I thought that they were saying that the atoms all stay in the same place physically and it's the position of the "holes", i.e. which molecule is lacking an electron, that move back and forth.
posted by XMLicious at 11:50 PM on December 7, 2008

I think you can get a better grasp of the significance of the memristor by looking at the figure in the middle of this page.

It shows the four electromagnetic circuit properties (voltage, current, charge and magnetic flux) and how they are related by the known circuit elements (resistor, capacitor, inductor). As a graph, the four property vertices can be connected by exactly six edges. All of these edges were accounted for except for one. Leon Chua in 1971 reasoned by symmetry that there should be a fourth circuit element, the memristor, in the lower right of the figure that related charge and flux to provide the sixth edge. This was just a mathematical curiosity until HP scientists actually figured out a way to construct such a device.
posted by JackFlash at 1:09 AM on December 8, 2008 [2 favorites has favorites]

Woodblock, this section of the wikipedia article says that using alternating current/voltage to probe the resistance will keep the value from changing too much.
posted by Pimonkey at 1:17 AM on December 8, 2008

*checks his memewrist0r*

Time to make the lolcats...
posted by Eideteker at 5:00 AM on December 8, 2008

I like how every technological advance will allow us to probe the workings of human consciousness. Flywheel governor automatically adjusts speed of steam engine? Sure, but it also may hold the key to human intelligence. Dish soap softens hands while you do dishes? Yes, but it may also finally unlock the ultimate understanding of the human mind. You know that little plastic thing on the ends of shoelaces? It may be the final piece of the puzzle of human existence.
posted by DU at 6:27 AM on December 8, 2008

But not to downplay this discovery. I'm sure it'll be awesome. It's the science writing I'm mocking, not the science.
posted by DU at 6:32 AM on December 8, 2008

Thank you for this!
posted by Nattie at 7:40 AM on December 8, 2008

I and my colleagues ... surprised the electronics community with a fascinating candidate for such a device.
That all changed on 1 May, when my group published the details of the memristor in Nature.
The story of the memristor is truly one for the history books.
In 1995, I was recruited to HP Labs to start up a fundamental research group
I was given carte blanche to pursue any topic we wanted.
Every once in a while I would idly pick up Chua’s paper, read it, and each time I understood the concepts a little more.

There's a reason scientific writing does not use the first person. It's because all scientists, deep inside, are insecure narcissits. When one of them actually does discover something interesting it comes spilling out in an unflatteringly immodest way. No idea if the science is good or bad, but the guy sounds like a douche.
posted by Nelson at 8:51 AM on December 8, 2008

George_Spiggott writes "How quickly can this be made to happen?"

Xezlec writes "Memristors suck. I mean they physically move ions around. Come on, that'll never be fast."

The article says switching happens faster than they can measure. That might be speaking to lack of measurement ability but still, pretty fast.
posted by Mitheral at 10:54 AM on December 8, 2008

What makes this revolutionary, even if its years away -- please forgive any oversimplification:

"A memristor is a two-terminal device whose resistance depends on the magnitude and polarity of the voltage applied to it and the length of time that voltage has been applied." So it's not just the voltage, but how long the voltage is applied. Thus this is not just a binary device -- it is an analog device, which is huge in real-time systems like bionics. But certain types of memristors change their state dramatically when charged, thus are suited as a binary device, which gives the potential for use as non-volatile ram.

It is read by applying either a very small voltage to it, so small that it doesn't alter the state, or with AC.

Neural synapses have a similar behavior - as the participating neurons fire more and more, the synapse reacts to this by changing to become more efficient. Sure the neuron action potential is binary, but the synapse is not. So it reflects the amount that it has be "charged", just as the memristor reflects the charge that it has experienced. This is what makes it perfect for mimicking the morphology and behavior of neural networks, making them simpler to produce.

Also, the application of memristors in crossbar configuration simplifies parallel computing, both in programming and scaling up to huge numbers of components.
posted by buzzv at 11:15 AM on December 8, 2008

I actually see an application of this outside the field of direct mathematics in computing. If I read the article (and the various interpretations here) correctly, then it sounds as if you could use memristors to construct the circuits of a robotic prosthetic limb, and shrink down the number of circuits you would need to do a particular task... say like "Raise armature, hold position" if you needed to lift a city bus...

err...

I for one wel....?
posted by Severian at 11:26 AM on December 8, 2008

A memristor capacitor is a two-terminal device whose resistance voltage depends on the magnitude and polarity of the voltage current applied to it and the length of time that voltage current has been applied." ... which gives the potential for use as is already used for non-volatile ram.

Interesting, yes

Major breakthough which fundamentally changes everything, no
posted by jpdoane at 11:45 AM on December 8, 2008

Nelson, this article was a memoir, not a research paper. I'm no EE but it sounds like this invention is pretty momentous, and it's OK to toot your own horn in such a case. Knowing how excited scientists get over even a small breakthrough after years of headbanging, I'd say this guy sounds pretty modest actually.

The invention of the transistor won a Nobel Prize for John Bardeen, William Shockley, and Walter Brattain. If the memristor does what the article predicts, I smell a Nobel Prize for this crew a few years down the road.

There really aren't too many ways to fundamentally change everything. But there are major inventions/discoveries that propel science and technology forward in large jumps, even if it's somewhat behind the scenes. Just for 2 examples closer to my own area, the discovery of DNA restriction enzymes and the invention of the DNA polymerase chain reaction (PCR) were significant enough to win Nobel Prizes, even though they didn't fundamentally change everything. They just changed enough to enable a lot of new research and technology. I'd settle for that.

After so many years of banging my head on so many different projects, I would like to have one turn out to be momentous. Just one. Is that too much to ask?
I guess it is.
*sniffle*
posted by Quietgal at 12:13 PM on December 8, 2008

This is marketing and doesn't look like any real revolution at all to me. Devices like this one have been presented for several years. Maybe this one has better on-off ratio than previous ones, but I didn't see them answer the most troubling questions: How fast is it and how many operations can it handle before it breaks down?

It doesn't seem to be a passive component either - an "ideal" memristor would be, but that comes from a strange definition of "passive" where the memristor just about squeezes in but the diode is left out. "Linear" and "Non-linear" are more relevant categories for components, and this one belongs to the non-linear ones.
posted by springload at 12:53 PM on December 8, 2008

Nelson: There's a reason scientific writing does not use the first person. It's because all scientists, deep inside, are insecure narcissits. When one of them actually does discover something interesting it comes spilling out in an unflatteringly immodest way. No idea if the science is good or bad, but the guy sounds like a douche."

Well, I wouldn't go so far as to make quite that broad a generalization, but in this case I think the analysis is correct. The guy does not sound like anyone I'd want to work for or near, fundamental discovery be damned. He gives his team an appallingly little amount of credit, and at one point basically castigates them for having too little faith in his discovery. Even if that's how things happened, the IEEE Spectrum is not the forum for it.
posted by Kadin2048 at 6:32 PM on December 8, 2008

I may have spoken too soon; I read the linked article very briefly, but now I had a look at the Nature paper also. This work is significant enough for that publication not due to its applicability, but because the technical definition of a passive component allows for some behavior that existing passives (capacitors, inductors and resistors) don't do, on their own or lumped together, that this memristor can do.

My understanding is that the memristor is non-linear and symmetric (acts the same whichever way you turn it), and that's it. This is what their measurement graph indicates. If that's the case it's very misleading to say that it takes 15 transistors to replicate one. My guess is that they made a computer model trying to reproduce the exact features of their device. That doesn't mean 15 transistors are needed to do its job.

There were also already devices that act like this (the Josephson junction is one, they name a few others in the paper). These other components have the charateristics of memristors, but the ways they function internally don't map all that nicely with the math. In the case of Josephson junctions, they also have to be cooled with liquid nitrogen to function at all, so they have less right than this memristor to be considered practical circuit elements in the same way resistors do.

Scientists like when sets are complete and this memristor completes, though in a somewhat forced sense, the set of passive components. Let me emphasize that this in itself doesn't mean anything about how we do electronics necessarily changes - we can easily make a two-terminal component that works the same way, but it's not a primitive then and needs to contain active parts.

Applications are a different matter altogether, and I doubt that this is the memory revolution coming. People have been trying with persistent conductance changes at various labs for quite some time as it's an obvious way to make dense non-volatile memory. From what they say in the popular article, the data retention seems good, though we don't know how well it scales. Speed doesn't seem to be a problem either, at least not for reading operations. One important parameter is left out though: The writing endurance.

My research group had a visitor a few years ago who presented some similar work. He showed some nice current-voltage curves and there was clearly a memory effect. When asked, however, he revealed that they could be written about 10-100 times before wearing out. These people bring up FPGAs as their killer application, which makes me think they have the same problem. FPGAs are typically written to once and then left alone forever.
posted by springload at 7:25 PM on December 8, 2008

FPGAs are typically written to once and then left alone forever.

This is not necessarily true; many FPGAs are actually configured / programmed after power-up using a bitstream from a (usually Flash) memory, and do not retain their configurations when powered down. But it is true that in many applications the configuration bitstream changes rarely or never, and it is true that some FPGAs have an onboard non-volatile configuration memory.
posted by musicinmybrain at 8:11 PM on December 8, 2008

This is not necessarily true

Not necessarily, but commonly enough that they can claim an improvement even with a chip that can only be written ten times, if there are other benefits such as speed or size. We can assume that such non-volatile configuration memory is what the memristor grid is to be used for.

They wouldn't talk about improving FPGAs if they could make broader claims of replacing flash altogether, and I think they are way behind in the number of writing cycles. It's a notorious and extremely difficult problem in developing new nonvolatile memory, and as far as I could see they don't address it.
posted by springload at 9:17 PM on December 8, 2008

I'm not an expert on this, but my boss is and he's skeptical about the importance of the memristor. This hysteresis effect in TiO2 has been known for many years, it's not a new observation, but it's a fitting of theory to explain something in experiment that is novel. Is it a new circuit element? nope.

I saw this guy talk a month or two ago and it was not clear this is any better than any other form of nonvolatile data storage either. He said they are currently implementing these in circuits so those papers will be very interesting, but I am of the opinion that it would not have been hard to include them in the first paper and would have made it a much better paper, so I am wondering what is taking so long...

So, in conclusion, as a researcher in the field of nanoelectronics, not everyone is convinced that this is an amazing discovery. That being said, our college is going to be having a memristance workshop soon, so a lot of people are interested in its potential.
posted by Large Marge at 11:20 PM on December 8, 2008

Paul Falstad's excellent Java circuit simulation applet was recently updated to include memristors, in case anybody's feeling too impatient to wait for them to show up at Radio Shack.
posted by teraflop at 4:54 PM on December 12, 2008