"What if the Future of Computing Isn't Silicon?"--Dr. Milan Delor, Columbia University

The revolution in artificial intelligence is sucking up a lot of electrical power, something that's growing at an alarming rate.  Science may have a solution in a new generation of highly efficient computer chips that use materials other than silicon and depend on light, rather than electricity to process data, bringing down power demands.  Dr. Milan Delor, a chemist from Columbia University, spoke about the groundbreaking work his lab is doing in a "Town Talk" sponsored by Telluride Science. His presentation was moderated by veteran broadcast journalists Judy Muller and George Lewis. 

Transcript

Science Straight Up

Season 6, Episode 8

“What if the Future of Computing Isn’t Silicon?”

Dr. Milan Delor—Columbia University

(Theme music, up and under)

JUDY: From Telluride Science, this is Science Straight Up

GEORGE: And this time around…

MILAN: People have realized in the past few years that, you know, the electricity demand for these server farms is really increasing to dramatic levels.

(sound-server farm) 

JUDY: Gigantic banks of computers—server farms---are going up all over the place as artificial intelligence takes off in a big way. Scientists worry about the consequences—the rapid growth in the consumption of power, water and other resources.

GEORGE: The folks in Telluride got to hear about that in a “Town Talk” staged by Telluride Science. The topic: “What if the Future of Computing Isn’t Silicon?”

JUDY:  Our speaker tonight. Dr Milan Delor is a chemist from Columbia University, where he leads the Delor lab. He's going to enlighten us about new materials that might be used in computing, basically posing the question in the title of his talk, what if the future of computing isn't silicon.” George, you used to cover Silicon Valley. What do you think?

GEORGE: Yeah, I think it's a possibility and a likelihood, too. I spent the latter part of my career at NBC covering the revolution in computer technology, and in 1993 the first year I started doing that, I got to interview Gordon Moore, one of the founders of Intel, a big chip company, and they were just about to release a chip called the Pentium, and it had 3.1 million transistors on a little piece of silicon. Okay, today, if you buy an Apple MacBook laptop with an m4 chip, it will have 28 billion transistors on a piece of silicon. And that number is probably going to go to 200 or more very soon, but we're reaching the end of Moore's law, where you can pack so many items on a piece of silicon, because we're now down at the nanometer level.

JUDY: So as circuits shrink, right? Yeah, sure, as the nanometer size, one billionth of a meter. Researchers are looking at just how much of the electrical electrical activity on those silicon chips is wasted in generating heat, and whether light in the form of photons rather than electrical signals in the form of electric. Trons might be used in future computers. With all that in mind, let’s welcome Dr. Milan Delor of Columbia University. (applause)

MILAN: Thank you very much, Today, I thought I'd talk about something that should concern all of us, and it's this idea that the demands for computing power that is increasing dramatically in the past few years, primarily because of artificial intelligence, is also giving rise to an exponential increase in electricity demand, and this is projected to continue over the next few years and to become really a big problem before 2030. So I'd like to talk about some potential technological solutions to this, and in particular focusing on using light instead of electrons to compute in a way that's much more efficient than what we're currently doing.

JUDY: Just think about all the helpful AI tools that are popping up on our screens: Chat GPT, Perplexity, Claude, Meta AI, Google AI, Microsoft Copilot. Every time we access them, we’re using a whole lot of energy.

MILAN: A typical modern-day user of Chat GPT, if you send a few queries back and forth, generate a couple of images, is the equivalent approximately of traveling 10 miles in a car.

GEORGE: Multiply that by the millions and millions of people using AI every day and the energy use skyrockets.

 MILAN: People have realized in the past few years that, you know, the electricity demand for these server farms is really increasing to dramatic levels.  About 5% of the total electricity consumption in the US goes towards these server farms. But the prediction, and from everything I'm reading right now, it seems to be an underestimate, is that by 2028 just three years from now, we'll be consuming about 12% of the total US electricity consumption just to power these server farms. If we extrapolate to 2030, server farms, will take approximately the same amount of electricity as the total industrial electricity demand today in the US. And if we look at what is driving this in terms of these predictions, is primarily for what they call AI special, specialized servers that primarily run things like Chat GPT.

GEORGE: The issue has made news.

JEFF TENCH, VANTAGE DATA: We suspect that the amount of demand that we’ll see from AI specific applications will be as much or more than we’ve seen historically from cloud computing.

JUDY: Jeff Tench, of Vantage Data, a company that builds those server farms, spoke about the power demands to Katie Tarasov of CNBC.

TARASOV: The A.I. frenzy has data center demand rising 15 to 20 percent every year between now and 2030. And as companies like Vantage build more and more, getting enough power is key. Natural gas is expected to fuel the majority of this and utilities will need to invest some 50 billion to support the growth.

TENCH:  A data center’s probably around 64 megawatts for the building itself.  And as we think about A.I. applications, those numbers can grow quite significantly beyond that into hundreds of megawatts.

TARASOV: 64 megawatts, 100 megawatts, how many homes on average can that power?

TENCH: Tens of thousands of homes per data center’s worth of power.

GEORGE: So, all those data centers consuming all that power, burning all that natural gas. But wait, there’s more. Many data centers use water to keep the chips in those servers cooled. With parts of the country running out of water, the problem is magnified.  Dr. Delor cited a Scientific American article arguing that AI needs to be more energy efficient, including a cartoon of a giant robot barbecuing planet Earth.

JUDY: As technology packs more and more computing power onto silicon chips, the copper conductors that connect the billions of transistors on those chips get tinier and tinier. And pushing electrons through those tiny connections generates heat because they can’t flow efficiently. 

MILAN: OK, and so the question that I want to ask is, what if we could make circuits that don't have these inefficiencies? So what if we could make circuits that have essentially zero resistance, where we could have our electrons or something else flowing past without any inefficiency, and this is basically the question that we ask in my research group at Columbia University. 

GEORGE: One of the materials they’re looking at is graphene, related to the graphite in pencils. Turns out, it’s a much more efficient electrical conductor than copper.

MILAN: We take Scotch Tape and you put it on this graphite, which comes from pencils or more pure sources of graphite, and you can peel off a single layer of this material. 

GEORGE: That’s right. He said Scotch Tape. A layer of graphene sticks to it pretty well. 

MILAN: And it turns out that when it's in a single layer form, it has very different properties to the bulk. In fact, graphene is the world's strongest material. It's stronger than a structural steel or diamond, and it's also the most conductive. And it turns out that there are also many other types of two-dimensional materials from other systems. And you can peel off one layer from one material and combine it with another layer from another material, and you make a new type of material that is better than the sum of its parts. And so we work a lot with this in my group.

JUDY: And while the scientists in his group are focused on superconductors like graphene, they’re also looking forward to a future where computer chips will run on light instead of electricity. Photons—light particles—doing the work that electrons now do.

MILAN: What we don't have currently is an optical, fully optical transistor. And so what I would want is to have light go through a structure, and to be able to come in with another laser and to switch that light off. This would allow me the on and off switching that is required for transistors. And if you think about what I'm asking for here, it's exactly like lightsabers in Star Wars.

(sound lightsabers)

GEORGE: Aah..those lightsaber duels in Star Wars.  And the moment when the two opponents lock their beams in a bind. 

(lightsaber sounds)

GEORGE: But because the real world doesn’t work like Star Wars, Dr. Delor says they’re still looking for a practical way to direct one beam of light with another, to switch those signals on and off and do computing. And one promising material is molybdenum disulfide. It’s actually sold in aerosol cans as a dry lubricant. It also interacts with light in interesting ways.

MILAN: We need to be able to steer light, and we need to be able to split light. Because we're using very powerful lasers to do this, but for technological applications, we will need to do this with a very weak laser, and we're not there yet.

JUDY: He said that when they DO get there, the speed of computing is likely to take a giant leap forward.

MILAN: The speed, actually, I haven't really talked about this, but the speed is enormous. This. We know already that we should be able to run algorithms 1000 to 10,000 times faster than electronics. The key challenge is to make light interact with light.  So with that, I'd like to thank you, and I'm happy to take any questions.

(applause)

JUDY: I was going to ask my favorite AI chat bot, Claude, some questions to ask you, but then I read that, in order for Claude to do that, or any chat bot, they generate heat in their silicon based servers, as you told us here tonight, and I was reading that that heat, in turn, requires a lot of water to cool the servers. So I'm very proud to say that I did not ask Claude and I saved water this week. Thank you. But what? When can we expect new materials like the graphene that you showed us, pencil lead, graphite, graphene to to allow computers to run at high speeds, but at the low temperatures? Is, Are we anywhere close?

MILAN: So in terms of the electronics part of computing, there are companies like Samsung that have invested heavily in two dimensional semiconductors like graphene, not quite graphene, but like molybdenum disulfide is the famous one, the one that I showed at the end. And they're serious about implementing two dimensional materials in their computing architectures, but I think it's a long way away. The reason that it's interesting, also that I didn't talk about is that it's extremely thin, right? It's atomically thin. And so you can start thinking about stacking things in three dimensions and packing a lot of these materials in a very small space, but in terms of years away, I'm thinking we're likely 10 to 20 years away from seeing a computer based on these types of transistors. But for optical computing, I think we're there. There are technologies there already.

GEORGE: In the near term, is the future of computing a combination of silicon and //this new generation of chips?

MILAN: These optical chips that I showed are commercially available. The silicon sort of industry is so advanced in terms of making these circuits on silicon, the most advanced photonic technologies are also in silicon. It's just that it's wave guides instead of electrons that are printed in silicon, and then it's light that travels through those silicon structures. So they're called silicon nano photonics, and so silicon is definitely going to be part of the future, despite the somewhat controversial title that we have. But that's not a bad thing. Silicon is widely available. Also the industry is very advanced. The foundries that make the silicon chips are extremely advanced. So we should benefit from that.

JUDY: So, we won’t be renaming Silicon Valley.

MILAN: No.

JUDY: Okay.

MILAN: But I think we will have, as you say, combinations of silicon with other materials, in particular for these optical computer chips. What I think is going to happen, and is already happening is that we put, you know, when we need these sorts of light-based interactions, we put other materials on top of the silicon wave guides. And that's that's likely going to be the future. And I would guess that this is already happening right now.

GEORGE: I'm fascinated by graphene because I remember when I was a kid playing with electronic circuits, you had to solder in resistors to some of the circuits, and those were carbon resistors. And I think of carbon as a material that offers resistance to the flow of electricity, and yet you've got graphene, which is a great conductor. How does that work?

MILAN: Graphene is very, very special. How does that work?  How am I going to answer…

GEORGE: I used Claude to generate an answer to that question. Let's see if it's true.

MILAN: Yeah.

GEORGE: that graphene has the carbon molecule, molecules arranged in little hexagons tied together. And when you do that, it becomes a great conductor. It doesn't become a resistor.

MILAN: Yeah. So that's that is true. So the structure of graphene are these hexagonal, so called honeycomb structures, honeycomb, all carbon, the way that you can think about it. So the way that I would answer it scientifically is that the the structure that defines how electrons move in that material is really quite special, and graphene is very unique. Actually, graphene is the only one that shows this, what we call a direct cone structure. The way to explain it intuitively is that if you look at your electrons in a normal semiconductor like silicon, so electrons, you may have heard, we often talk about light being both a particle or a wave, and similarly, electrons are also both a particle and a wave. And when we talk about electrons in graphene, we often talk about a wave, and that wave, the wavelength of that wave is very, very large, and that allows electrons to feel very little resistance, essentially, so they flow very much like a wave through graphene, and that is what makes graphene so special.

JUDY: Are there ethical considerations or potential impacts of your research that you find important to consider? Or is that ever a consideration?

MILAN Ethical considerations. 

JUDY: Are there any? 

MILAN: So the students in my group would say, sometimes, some students in the group might say things like, you know, we should be very careful about where we get our materials from and where we get funding from. And so those are more the ethical considerations that we think about. 

JUDY: Not worried about lightsabers? 

MILAN: No. Lightsabers, no. And you know, the one, the obvious one, I think, is people who work in artificial intelligence, which I don't consider myself working in, would have you know, there are ethical considerations with the interaction with society that are important. But you know, we are very much on the fundamental side of research. 

JUDY: I think that horse has left the barn.

GEORGE: You describe this kind of exponential rise in the need for energy with AI, what is the sense of urgency to solve that problem. Are you feeling that in terms of pressure from outside?

MILAN: Personally, no, but yes, when I read that report, actually, I was really shocked to see how bad it was. And I think there is a collective conscience that's building among scientists to say, okay, this is a big deal, and we need to address this, you know. So there is going to be probably, in the near term, the biggest challenge that's going to happen is simply in the architecture, the software architecture to run AI right, which can be more efficient than currently, and many people are working on this already, and then after that, the hardware changes will come, I think. But there is, there is a lot of pressure.

JUDY: You mentioned funding, and these two successful companies that are already involved, is funding really a problem for something like this, because it seems a lot of funders would want to be invested in doing this.

MILAN:  I think it doesn't seem to be a problem right now. I think there's a lot of money going into this right now. Yeah, not so much for fundamental research. I wish there was more money going into academia, like (laughter from audience)

JUDY: The entire side of the audience, basic research people, I'm going to guess.

GEORGE: it seems to be a recurring theme week after week here,

JUDY: Yeah, okay, we'd like to open it up to the audience. 

GEORGE: Mark Kozak, CEO and Executive Director of Telluride Science, had the first question.

MARK KOZAK: Does graphene occur naturally, or is it manufactured?

MILAN: A lot of these materials occur naturally. Yeah. So not the two-dimensional form, but in both form, they occur naturally, and then you can take one layer off. But the graphene that we would use in technologies is not naturally occurring. This is synthesized in the lab on large scales. It's very cheap to do.

MARK: Okay, thanks.

MAN IN AUDIENCE: How do you generate the light that you're using in the chips? And doesn't that isn't that in itself, going to give off heat?

MILAN: Yeah, that's a great question. So absolutely. So the main energy consumption actually in these optical chips is going to be how we generate light. And the way that is normally done is these is on the chip as basically an LED. So like most of the, probably most of the lighting in this room, it's the same type of technology, and that generates light is very efficient, but you're right that it consumes electricity and it also generates heat compared to what we're doing with these transistors. However, it's very low because, in principle, what will happen is that you generate light in one location, and then you split that light in multiple, multiple routes, and then you use all of these different split circuits. And so you've only had to generate it once. So the energy efficiency estimated for these chips for the same type of computing power is about 100 times better than electronic based chips.

JUDY: Yes, back here, yes.

MAN IN AUDIENCE: So biological systems. There are biological systems that convert light, can transduce light into electron transport or motion, even. And so I just wonder, and maybe this is naive, could one Imagine designing or using proteins that would be able to be transistors in circuits like in computing?

MILAN: Yeah, I believe there are. I'm not specialists at all, but I believe this is a technology that people are exploring as DNA computing, for example, where you use essentially biological or biologically inspired systems to compute. Another one is neuromorphic computing, which is trying to sort of work like our brains do, and that's definitely a very active area of research that I am not at all familiar with.

JUDY: A young man.

BOY IN AUDIENCE: What do I need to do to help solve the problems? I'm a middle school. School.

MILAN: What do you need to do to solve the problems that we discussed?

MARK: Yeah, what does he need to do to help solve the problem? 

MILAN: Come to Columbia University.  (laughter from George and audience)

JUDY: He’s in middle school.

MILAN: Come to my lab. 

MARK: He’s offering you a job. (laughter)

JUDY: I have a question for you, and we often ask scientists this, and we've forgotten recently. But can you remember a moment when you were a child, when you got hooked on science?

MILAN: I grew up on a small island in the southern. Indian Ocean is Reunion Island. It's a French island, but it's really a kind of nice place to grow up in. It's also a very extreme Island. There's like, very active volcano, so I saw like two eruptions when I was a kid, and a it's a very famous surf spot. There are giant waves, and everyone scuba dives, and so you're really exposed to these elements of nature. My best guess is that this is what drew me to science to try to understand what was going on. But also, I don't know, sometimes it's just a lot of random decisions that end up guiding you somewhere that's probably the more the less romantic but more correct answer.

JUDY: No eureka when the volcano erupted.

MILAN: No.

JUDY: Do we have any more? 

GEORGE: I have a question for you, Judy.

JUDY: Uh oh (laughter)

GEORGE: Why is a superconductor like a silent Buddhist retreat?

JUDY: Why is a superconductor like a silent Buddhist retreat…(pause)

I don’t know, George.

GEORGE: No ohms. (laughter)

JUDY: Don’t laugh at that!  Do not encourage that!  Okay, thank you all for coming,  let’s give a big hand…(applause)

(sneak in theme music)

GEORGE: We want to thank Dr. Milan Delor of Columbia University for his enlightening talk, and to Telluride Science for making all this possible. And a big thanks to the Telluride Mountain Village Owners’ Association. Dr Delor appeared before a live audience at the Telluride Mountain Village Conference Center and our audio engineer was Colin Casanova.

JUDY: Mark Kozak is CEO and Executive Director of Telluride Science and Cindy Fusting is CFO and managing director. Sara Friedberg is Lodging and Operations Manager and Annie Carlson is Director of Donor Relations.

GEORGE: If YOU want to donate to the cause, go to Telluride Science-dot-ORG. That’s also where you can find our podcasts. And on your podcast apps like Spotify or Apple, look for “Science Straight up.” I’m George Lewis.

JUDY: And I’m Judy Muller, inviting you to join us next time on “Science Straight Up.”

(Music up full and then fade out)