The Tech Elon Has Been Waiting For

Opening

0:00Industrial FA flash memory tops out at around 200° C and above that there's essentially nothing that works. Okay, it's a memory which is a memory switching device and it's made from graphine halfium oxide and tungsten and it operates reliably at 700° C. Really cool. I mean it's hotter than molten aluminum. Aluminum melts at 660°. If you put this thing in a vat of aluminum, it will still retain memory. Hello internet. This is your captain

Intro

0:31speaking Lester Narre joined as always by my co-host and our resident PhD Krishna Chowdery. We have a special deep dive today for you all. Be sure to check out our rundown from last week. But this week, we'll be sure to heat up the future of electronics as we look at a breakthrough in high temperature memory. This is something that uh Elon has been waiting for for a long time that may enable electronics that can survive lava. I'm going to need the details. >> I'm going to need to understand what's going on here. >> This comes from a new research paper in the Journal of Science. Shout out to

1:12science from March 26, 2026. It was a collaboration between researchers at USC here in California, the Air Force Research Lab in Dayton, Ohio, and the Kumamoto University in Japan. Please correct my pronunciation with funding for this research being provided by National Science Foundation, the Army Research Office, and the Air Force Research Lab. So clearly they think there is military applications. I'm sure it relates to hypersonics. As always, we are going to learn about the science from the ground up today because this is from first principles.

The heat limit of modern electronics

1:51[music]

1:56[music]

2:02[music] Today we're covering an invention that I think is going to be a gamecher for the future of computing. You know, we live in the digital age of electronics. Chips are the bedrock of our economy. And yet, all of those chips really have a Goldilocks zone when it comes to when they work. Goldilocks in the traditional sense of like not too hot, not too cold. The this research is getting to that not too hot part. You know, the Papa Bear that uh that was like too hot, the porridge was too hot. Now, we're getting a chip that can withstand that porridge. Okay. Um, these are memory chips that happen in

2:43our phone, in our laptop, in the cameras that we have here in the studio. They're all designed to work reliably roughly between 85° C, so 200 Fahrenheit for our Americans to about 150° C, so 300 Fahrenheit. Okay, that's where they sit. and industrial fat flash memory tops out at around 200° C even with a lot of the heroic packaging where you try to insulate temperature and things like that and above that there's essentially nothing that works

Why extreme-temperature computing matters

3:18okay there's no computation happening above 300° C but that doesn't mean that we don't want to have computation happening at that temperature okay there are a lot of applications that even I can think of here not being in the field. That would be really great to have at 300° or above. Okay, for example, oil and gas. Oil and gas drillers, they push deep into the Earth's crust. The deeper you get into the Earth's crust, the closer you're getting to the mantle where all of the magma is. So, the temperature is going to go up. At 5 km, you're flirting with like 150° C that's already pushing it. At 10 km, you're well past 300°. And the

3:59sensor packages that we have today, when you send them down these bore holes, they need to either operate with like an active cooling thing where like there's something with the the drill that's trying to cool the sensors. That's always >> not convenient. Not ideal. >> Yeah. And then and then they have to transmit their data to the surface in real time. They can't do any computation down there. >> Interesting. >> So then there's this like massive wiring problem that you have to deal with. The wires are going through this really hot. >> The engineering is way more complex unnecessarily if you could just have it. >> Yes. If you could just have like a chip that's right on top of the drill, right?

4:40That's right there. Then you can have something like an autonomous downhole drill like electronics that decides where to go right there. I mean, we've got all this AI, but we can't actually put that to use when we're going all the way down there, right? If in I'm going

Drilling, hypersonics, space, and Venus

4:56to make another Armageddon reference when we need to go to an asteroid and and and we need to draw I guess that's a bad example because it's not that hot. >> Well, I mean, but even the friction is going to create some hot >> some heat. You know, we're going to want to be able to do that. >> Yeah, we just >> planetary defense. >> Yes, as always, you [laughter] know, we are defense first here, not war, only defense. Now, you mentioned hypersonics. That's another big thing, right? Um, in aerospace engineering, the inside of a jet engine or the inside of a rocket engine or the nose of a hypersonic aircraft, these things are getting to really, really high temperatures. We've covered this in previous stories where the the temperature gets so high that

5:38chemistry starts mattering. >> It's a little weird. >> Yeah. And it starts getting a little weird. And we can't make any intelligent decisions on site because of the chip. The chip is just going to get fried. And finally, the thing that Elon has been worried about and all of these tech bros are worried about is putting data centers in space, right? >> This is the number one conversation in SF right now. >> Yeah. Yeah. I mean, Elon wants to put data centers in space. He's not the only one. Um SpaceX had an IPO of $1.7 trillion. And I think a lot of that hinges on Elon's promise that he can put AI data centers in space. We're going to get into the details there, but there's been a lot of push back because of

6:21fundamental physics, right? They're saying, "Oh, like fundamental physics is telling you that you can't put data centers in space." Well, yes and no. Okay. Yes, in the sense that yes, the physics and the electronics that we have today are it's probably not a good idea to put data centers in space. But with this problem, you can have innovation like the one that we're going to talk about. And perhaps it's possible. The main problem really is that to keep a 1 megawatt artificial intelligence server cool in the vacuum of space, you need a radiator that's roughly the size of four tennis courts, right? That's because the space is a vacuum. There's no way to extract heat other than radiation. And

7:04and radiation is not a great way to let go of heat because you're relying on photons. It's not is it would you It's like because it's not efficient. >> Yes, exactly. It's just not efficient. We're going to get into some of the physics of that later. Um and the my favorite sort of conundrum that I think this is going to solve >> is planetary exploration. I think you're going to like this, too. >> Um there's a reason why we have so many photos of Mars and not that many photos of Venus. Okay. Mars is pretty chill to go to. It's like kind of cold. >> Yeah. Literally. It's like literally it's farther away from the sun and it's fine. Venus is closer to the Earth and it's got this massive greenhouse effect

7:45which means that it's really hard to get put stuff down there. Actually, the the first guys to do it were the Soviets. Venus has a surface temperature of about 465° C and an atmospheric pressure that's 92 times that of the Earth. The Soviets put down the Soviet Vanera landers and they're the only spacecraft that have ever survived on the Venus surface. >> Guess how long they survived. Over under an hour. >> Uh it's got to be under an hour. >> It's under an hour. Yeah, it's it's between like 23 minutes. I think the record was like 2 hours later on, but like >> 2 hours. Yeah. Yeah.

8:25>> Right. Like the the stuff that JPL puts on Mars has lasted like 20 years. >> Yeah. It's two, we're talking about two decades versus >> versus two hours, >> right? Right. Two hours is the best that they can do. And >> I will note this is for those who are TV fans. >> Uh we've talked about For All Mankind on the show before, but Apple has a second Apple TV has a second show called Star City that looks at the Russian space program in the same alternative history timeline where they got to the moon first and they do a whole all of season 1. Part of it relates to the Venus mission. >> Really, dude? I got to watch that. >> And it's it's it's alternative reality. So, >> no spoilers.

9:05>> I I love that kind of stuff. >> But it it is interesting to kind of understand the technical challenges which they brushed on at a high level in the show. But recently, I've been watching it and this brought it up as a concept and the engineering problem is non-trivial. >> Yeah. No, completely right. It's like going towards the center of the earth. It's the same drilling thing. High pressure, high temperature. your electronics are going to get completely wrecked. Right. The photo we have from the Veneera 9 in 1975, this is the best photo they got. It's the first image of a planet other than the Earth. Credit to the Soviets. Yeah, they did. >> The first image of a planet other than the Earth was not the United States and Mars. It was Venus and that's all they

9:47got. But still, that's kind of crazy. >> But who got to the moon, though? >> Yeah. That's that's I mean, they were like, "Okay, let's go let's go the other direction." >> Yeah. Yeah. Yeah, >> the race is actually this way, guys. >> Yeah. Um later missions, um they were actually able to get some color photos. This is the one that lasted 2 hours. >> Okay. >> And you could get some color photos. You can see the sort of yellow haze that we know that the Venus has if you point a telescope to it. But again, only 2 hours. And you can see there's like it it kind of looks like the surface of a volcanic debris field because there's a bunch of volcanoes on Venus. Um but

The 700°C memory breakthrough

10:23that's that's it. That's all we could get. And we can imagine, you know, as we continue to expand in our own local solar system and otherwise, having electronics that can survive high temperature environments is going to become >> Yeah. >> a valuable asset to expand where we can actually go. >> Exactly. I mean, I'd like to know more about the surface of Venus, right? We know we we have better surface images of Mars, Pluto. We have we have better surface images of Pluto than we do with large swats of Venus, which is our nearest neighbor. Venus is closer to us than the Earth. And we still don't know what's under those clouds. >> So unbelievable, >> right? And it's just because it's just it's it's a completely

11:05messed up situation to get down there on the surface. Everything fails. >> We can't stand the heat, so we can't go in the kitchen. >> Yeah. Exactly. Exactly. Um, and into this landscape comes this paper that was published by USC. It's describing something I think genuinely new. Okay, it's a Meister, which is a memory switching device. We're going to get into what that is. And it's made from graphine, halfium oxide, and tungsten. And it operates reliably at 700° C. >> That's unreal. >> 700°. Yeah. They've they've raised the temperature like by a factor of more than two. And so there's something fundamental. There's a there's a real breakthrough

11:47here that's beyond an iterative upgrade. >> Yes. Yeah. There's it's it's I think it's really cool. I mean, it's hotter than molten aluminum. Aluminum melts at 660° C. If you put this thing in a vat of aluminum, it will still retain memory. It's hotter than the surface of a lava flow on Earth. Not all lava flows. I know there's going to be people in the comments being like, "No, >> the vulcanologist." >> Yeah. Yeah. The volcanologist is going to be like, "No, lava is hotter than that." Actually, um, there is a volcano in Tanzania, [laughter] the old Doyo Langai volcano, where the lava is 600° C. Okay. So, there exists a lava on Earth where I could put this

12:27chip and it would still work fine. So, the tagline still kind of works, [laughter] right? If if you if you took it to Hawaii and put it in the lava in Hawaii, it would not. I think Hawaii's is like a thousand plus. But it perhaps might and we'll get into that later. >> We're on the path. Yeah. Um the other crazy thing about this is the chip the the Meister device it maintained stable data retention for 50 hours. Really >> 50 hours. And it retained that data retention after a billion switching cycles without failure. >> So this is both high temperature and long duration comparatively to where we

13:08were before. Yeah. And the physics is actually like it gives you both >> which is in cuz both are different. >> It's like a graph on different axes and you can have high temperature. It's like okay great but it's a microcond. >> Yeah. Yeah. But and here what's what's cool is it's the same physics that's giving you both endurance limits which I think is really really cool. It's by I think it's by a large margin the highest temperature at which any nonvolatile memory has ever been demonstrated to work. >> This is it's really really cool. And I mean why should why should like normal people care okay not scientists? Well I think the border between environments where we can put intelligent electronics and environments where we cannot is

13:49about to shift >> dramatically >> and then our understanding of everything about the world around us changes in a in a very big way. >> Yeah. I mean it affects how we drill energy, how we explore planets, how we monitor volcanoes, how we put data centers in space. That has a lot of social and political implications for our society because then maybe maybe we don't need to, you know, like take everyone's water. >> If you don't want a data center in your backyard, this story matters a lot because we can just put them in space. >> Yeah. Yeah. I mean, it's it's one of these like quiet material science results that uh it doesn't make the front page of the New York Times, but I think it's going to expand the frontier

14:30of what our civilization can do. >> You're saying there's a chance. >> Yeah. And I think I think I think it's an incredible story. We are going to get into the details of I this has been such a big conversation in the business and technology world. we are so overleveraged on AI in the US uh in terms of its impact on our GDP. Everyone's saying it's fine because we're going to put data centers in space and it's been very contentious as a

Why conventional electronics fail in extreme heat

15:02one aspect of that larger conversation. This the potential for this to be an enabling layer is interesting. But before we get into the details of that story, a brief note of housekeeping for our longtime listeners in FFP Nation. Welcome back. If you happen to be catching this episode for the first time, it is the two of us here bringing you the latest breaking science news every week and any way that you can help to support the show so we can combat against the billionaire algorithms. Like, a share, a comment, bring it to Journal Club, put it into the group chat. All of those things help us reach more people with the way in which our

15:44world is changing and growing because of the advances we have at the frontier of science. Our website has all of our episodes, all of the research papers and show notes that we cover. So, if you are someone who's a grad student, a posttock, or someone who just really wants to get actually more into the weeds of any of these stories, go to ffpod.com. You can check out our research paper library. If you want to support the show, you can donate at ffpod.com/donate. And you can also check out the merch shop if you want to let the world know that you yourself are a member of FFP

16:25Nation. But enough of the riff wrath. Let's get back to this very juicy story. I do like my steaks medium rare, [laughter] but I I guess I can make it much quicker if I do it at 700° C. >> Yes. Yeah. It would it would only cook the outside like micron [laughter] >> and then it would be all rare. Um Gordon Ramsay would love that. So, let's first talk about why the conventional electronics that are in our phone, that are in the laptop, everywhere else, why they fail at high temperature. >> Why I can't bring my iPhone and put it in a lava flow. >> Yeah. Yeah. Yeah. I'd like to. >> Yeah, I could put it in water. >> Yeah, but but lava lava is something you

17:06can't do. Okay. [snorts] The physics of high temperature failure. We first have to understand the traditional CMOS architecture that is at the base of all of our technology. This is complimentary metal oxide semiconductor. These devices fail at around 200° C. Here's how it usually works. Okay, so at 0 Kelvin, which means absolute zero, everything is everything is stationary. Silicon is an is a insulator. What ends up happening is atoms have something called a conduction and a veence electron shell. Okay? They've got parts where veence electrons can occupy

17:46all of the shell. So there's nothing going on. All of the shells are occupied, right? Then there's a conduction shell. There's a conduction energy level where there's like one or two electrons hanging out. And those electrons are the ones that skip from one atom to the next to the next and they create current which is what we use. Okay. Now at 0 degrees Kelvin all of silicon's veance band is completely filled. If you put a bunch of metal and a bunch of silicon together the gap between the conduction band and the veance band gets bigger just because of neighboring effects. It's like it's like the the neighboring atoms start pulling

18:28on your own energy gaps and it expands the gap between these two energy bands. >> The wiggle c starts getting more intense. >> Exactly. And so at finite temperature, what you can do is you can bump an electron. You get some wiggle, right? Because like there's jiggling happening. There's photons that are at some energy level that's going to come in, strike an electron. the electron is going to move up to the conduction band and then voila, you can have a little bit of conduction. Okay, now let's get into exactly how this happens. We've got a video that shows exactly how this conduction and veance band works. So on the bottom you've got your silicon. You've got all of the sites that are now

19:08occupied by electrons which are in the >> in the blue. >> It's like the housing market. There's nothing on the market. >> Yeah. Yeah. Yeah. There's no openings. Now all of a sudden somebody sells a house. So an electron from the veence band moves into the conduction band. Now that conduction band electron can move around and create a current. Notice at the same time the hole that was left in the veance band can also move around. It's like you have a bunch of packed theater seats and everybody moves one seat over >> but then the the the empty seat >> can can start moving around too. That's a positive charge in some sense, right? because it's a hole. The electron has

19:49gone away. So, it's left a positive hole. And that hole can move around in the same sense that the electron can also move around because >> because it's attracting because it's that empty space and it needs to be filled and that creates the movement. >> Yeah. The movement, the actual movement, the physical movement that's happening is always electrons, right? It's the negative charge carriers. But the the the the sort of thing that matters in the veance band is that positive hole, >> right, >> that's moving around, right? It it unlocks the stationary state. >> Yeah. >> Uh that exists because it's now moved up to >> Yeah. Yeah. Yeah. Exactly. And so you can have two different types of semiconductors here. You can engineers can make silicon switches using either

20:30phosphorus or boron. You can dope the silicon with either phosphorus or boron. And then these artificially dictate how many charge carriers you're going to get. Are you going to get more electrons in the conduction band? Are you going to get more holes? So that's where you get if you've ever heard of like pt type and n type semiconductors the n type is the negative type semiconductor because the electrons are the ones that are moving around the ptype is the positive type semiconductor because the holes are the one that are moving around and when you combine the two you get things like LEDs transistors but at too high of a temperature you're going to get wrecked. Why? Because if you have too high of a temperature, everything is going to go into the conduction band because there's so much

21:11jiggle and then I'm just going to get a short circuit. It's always going to be conducting. With a transistor, I'd like to control when it's conducting current for a one and when it stops conducting current and becomes an insulator for a zero. But at a high enough temperature, it's always going to be conducting cuz there's so much jiggle that the electrons are just like, "Oh, I'm just going to I'm free to move around. >> Everything moves." Okay? And so part of what we're saying is we're trying to engineer a level of control >> of how this movement happens. And at high temperatures currently in our traditional uh chips and systems, they are unable at high temperatures, you lose the ability to engineer and control the jiggle, the movement, how things

21:51move to the conduction layer and back according. >> Yeah. Yeah. You can't control the transistor going from a one to a zero. It's just always going to be a one. >> Right. Right. Because it's always >> Yeah. And that's the the bedrock of computation. Right. Right. Um now people have managed to make the band gap way bigger because if you make the band gap way bigger then the temperature might not be enough to actually put it up. Right. >> Right. And then and then you can like preserve this transistor property. But and these are authors the authors of the paper that we're talking about they cite this research from NASA research. Um, it shows that you can you can get all the way up to 800 degrees C, but memory is still an issue. Okay, you've created a

22:33sort of circuit element here, digital integrated circuit, but nonvolatile memory, something that retains data even when it's powered down, right? The the stuff that's in our RAM, like I I turn off my computer, it's not like I'm going to lose >> everything >> everything, right? That kind of stuff is still not up to that scale. So through public funding we discovered that it's possible to go above 800. Uh but the the issue still remains memory does not in this context have that high temperature limits. >> Exactly. And if we want to do anything we got to store data right we got to store weights. If we want to do a neural network we got to store what the algorithm is supposed to do. if even if

23:14without the AI stuff, right? If it's just like instructions on what to do, that is stored in your RAM and if the RAM fails, but like your your circuit is doing fine, what's it going to do? There's no instructions. Right. >> Right. Um so most of the applications that we have like in our computer in in the phone, they use something called flash memory. This also uses a transistor. This is the MOSFET transistor. um where you combine a P and an N type and it runs into the same issue with the high temperature because you're using P andN types, right? Um and if even at like a low temperature like room temperature, right, consumer

Flash memory and resistive RAM

23:54electronics like the flash in our in our computers, it wears off after about 10,000 to 100,000 readwrite cycles, >> right? So they they're not they can't run forever. Yeah. There there's deterioration. Doing work deteriorates just like our own human body. Yeah. >> As an athlete, you can't be an athlete forever. Your body eventually gives up. Similarly, in this context, flash memory deteriorates at some time interval. >> Exactly. Use. >> Exactly. And so because of that, people have been slowly looking at other substrates of computation, specifically something called resistive memory, RAM. Okay. Um, this is resistive random

The missing fourth circuit element

24:36access memory. And to understand that, we really have to understand something called the memister. Okay. Okay. Let's go back to 1970. There are three types of fundamental circuit components. The resistor, the inductor, and the capacitor. This is the stuff of undergraduate physics. And if you're an electrical engineer, I'm really sorry for your loss, but you probably you probably deal with this every single day, right? Um, a resistor is something that resists electric current. So, it drops a voltage across a resistor. And um, an inductor is something that takes a changing current and creates um,

25:17another current that's going in the opposite direction. So, it kind of stores energy using a change in current. And a capacitor is something that stores energy using charge. Okay. Now, Leon Chua in 1971, he was at UC Berkeley, he said, let's look at these three components that we've got. We've got a capacitor, right? A capacitor is something that has a relationship between charge and voltage. If I store a bunch of charge, that stores a voltage difference. I've got a resistor. That's a relationship between voltage and current. Right? As current goes through, there's a voltage drop. And I've got an inductor. That's

25:58something that has to do with current and magnetic flux. Okay? There's four variables, though, >> right? >> And there's only three components, right? So, for those who may be audio listeners, we're looking at a square. It kind of looks like those uh those one, two, three, four. I don't know what you call. >> Oh, the Yeah. Yeah, the thing we used to do in elementary school that would tell me like what my fate was. >> Yeah, the paper thing. Yeah, >> you can let us know in the comments. But but on each of the four sides you have exactly the components you mentioned and each of the four corners you have the charge, voltage, current and flux. And

26:38the point here being we've talked about every relationship but one side of the square >> but one side of the square right there should be Leon Chua decided there should be a relationship between flux and charge the flux capacitor >> in some sense. Yeah. Yeah. And he deemed this thing the memortor he said there should be something out there. Okay. This is a component where the resistance changes dynamically based on the charge flow. and he he ruled that if there is such a relationship what this thing is going to do is it's going to remember the state when the power is off. Okay, let's the telltale sign. He he also showed okay if if experimenters are

27:19looking at this thing, he was a theorist and he said if experimenters are looking at this thing, all you got to do is look at the IV curve. Okay. The um the IV curve is the current and the voltage >> on the two axis. >> Voltage on the axis. >> Yes. Okay. And if we look at the IV curves, they're all very unique. On the top left, that's a normal resistor. Okay. As I increase the voltage, I'm going to increase the current across the resistor. And there you've got a line. This is a nonlinear resistor. A linear resistor would be like the stuff that we do in undergrad physics is just like, oh, V equals IR. That's not entirely true, right? The R is a function of the

28:00V. So it there's going to be like some nonlinear resistance where like at low voltage the resistance is not that high and then at high voltage the resistance is way higher. Things like that. Um or maybe no it's the other way around. At high voltage the resistance is lower because you know there's more pull. So like the it's easier to shuttle electrons through. Okay. So that's what we're seeing in the top left corner. Okay. On the top right corner and the top in the bottom left corner we're seeing the IV curves for a capacitor and an inductor. Okay. Those things cycle around those things. And this is something that everyone knows from, you know, if you've ever done RC circuits, you'll you'll know that there's an

28:41oscillation that happens in RC and RL circuits. Especially if you if you combine a capacitor and an inductor, the capacitor gets charged and the inductor doesn't get charged and then the inductor gets charged, the capacitor gets discharged. So there's a cycling that happens and that is something that is characteristic of like s and cosine. You got the unit circle, the x-axis is s, the y-axis is cosine. Great. Leona said, "I should be looking for something called a memortor where charge and magnetic flux are related to one another. And what I would see is a curve that looks like a loop, a figure 8 loop.

29:24And the challenge is there is no way to combine the other three curves to get that fourth curve. >> Interesting. >> Okay. It is an independent element in electronics. >> It's it's fundamental. It's not derivative. >> Yes. Exactly. You know what? You see what I'm saying? It's like [clears throat] it's like you can't make you can't make an electron out of quirks. >> Right. >> Right. Similarly, you can't make a memory by just cleverly placing resistors, capacitors, and inductors together. This is an independent component that is one of the four fundamental components of electronics. >> And this sort of goes back to our previous visual, which you know in this

30:05context of those four parts, you needed this is sort of the missing piece to the puzzle of the fundamental elements. >> Yeah. And it's just a mathematical argument >> which is so >> which is so cool to think about right that it's just like there's there's four variables we've only got three components there should be a fourth one theorists have value everybody okay >> it's it's very cool and one more thing I want to say about the memor IV curve right um notice that for a capacitor and an inductor at zero voltage there is a nonzero current okay that means it can't store data Because if I turn the thing off, there's going to be current and

30:46that's going to dissipate whatever the the stuff was happening, right? A resistor could store data because at a zero voltage, there's zero current, but it only stores a one. It's whatever the resistance is. So, it's not really storing anything. If we go back to Claude Shannon's um information theory, I need two things to store data. I need a zero and a one. >> A memory, on the other hand, at zero voltage, it can be in two states. It can be in a low resistance state or a high resistance state. >> This is so good. This is so good. >> Now all of a sudden people start paying attention because this could be a substrate for nonvolatile memory. Meaning I turned the thing off. But

31:26depending on the history of where I was on the curve, I could be in either a low resistance state or a high resistance state. If I can sense that, that's my 0 and one computers. >> As a quick uh visual analogy here, cuz we're basically looking at this figure 8 that's kind of like a racetrack. You remember Hot Wheels as a kid? >> Yeah. >> So, you remember how you could have the you could have a figure eight like that, but one side would be going under and then it would come back and it would be going over. And that's kind of what we're saying where it can be in two states at that center point. It can be on the top part of the Hot Wheels track or on the bottom part of the Hot Wheels track. And that's that mechanism is how we're storing the zero and one. How we're able to have two states, which means we can have a zero and we can have a one. >> Exactly. Yeah. And for the Formula 1

32:06fans, the only circuit that does this is Suzuka in Japan. It's the there's a figure 8 that's beautiful. One day we'll go. Um, so the theorist has done his job. He's like, there should be this thing and specifically this is what you guys should look for. >> Um, it takes 40 years.

The memristor is finally discovered

32:26>> Oh my god. Nearly 40 years. Stanley Williams at um, Hullet Packard Labs. So that's HP Labs. I love these nature articles that are just like bombshell articles because the title is just a banger every single time. The missing memor found after it's nearly 40 years after Leon Chua postulated that this fourth passive element should exist. >> Yeah. Um Stanley Williams announces it in nature and he shows that there is a meister that he's created using titanium dioxide and platinum electrodes and it uses something called oxygen vacancy migration. >> I just want to note that this we're

33:07talking about 2008. >> Yeah. >> We were we were in high school. >> Mhm. >> Uh so it wasn't Isn't that crazy? >> It wasn't that long ago. >> It wasn't that long ago. No, we were in high school when the first memeister like when it went from theory to reality, >> right? ex and it's still at a research uh stage level. We're not talking about in end consumer products. >> Yeah. >> The iPhone came out a year before this. >> Yeah, that's crazy. >> Just to just to put it in time context. >> Wow. >> Which is >> that's a good way to Yeah, that's a good way to think about it. Yeah. I mean, we had an iPhone before we even knew that a memor was possible, [laughter] >> which is so crazy

How oxygen vacancies store data

33:45>> knowing knowing what we use it for now. >> Yeah. Yeah. I mean, it's insane. Okay. So, let's let's talk about this veilance charge mechanism that I was telling you about about how the memberister works. Here's the idea. Um, so I told you, right? He he got a titan titanium electrode, a platinum electrode, and in the middle were oxygen vacancies. Now, vacancies is something that I alluded to earlier when I was talking about the holes. The holes are vacancies in the sense that I have taken something out. But now those those holes where I took that thing out, those are mobile just like the the the stuff that I took out. Those are mobile elements. And if those mobile elements can rearrange, they can create

34:26electronic components on their own, right? In the same way that a ptype >> semiconductor can use the holes to do all of its electronics. The point being that vacancy is not uh stationary. It has the ability independent of the thing that left to move on its own. >> Yeah. Exactly. >> Okay. So, um let's get into a little bit more about how these oxygen vacancies work because this is going to be the central thing we need to understand about why the physics of this modern paper is working. Okay. Um when you have these two electrodes, you've got a positive electrode and a negative electrode, right? You you what you do is

35:08a switching process. Okay? So this is effectively how it works. Let's say I've got a top electrode and a bottom electrode. In this case, it's like silver and um titanium on on the bottom. But it doesn't really matter what the identity of the two electrodes are yet. It's going to matter later. What happens is in order to understand what this veilance charge mechanism is let's let's try to think of a analogy a simple analogy okay imagine like a really thick densely packed forest like the forests of Endor in the Star Wars episode 6 return of the Jedi. >> Okay. >> Okay. >> Okay. >> Now if you remember the forest of Endor

35:51that they had those like motorbike thingies. Um I actually looked up what they were called. that were called forest motorbikes. 74Z speeder bikes. Okay, remember like they they they like zoom through the forest at really really high speed. Now imagine a really really dense forest. Okay, if you've got a speeder bike in a really really dense forest, you're going to crash. >> Yeah. You have to be an uh Lewis Hamilton. >> Yes. Yes. And even then, you know, like like I'm talking about like there's there's a tree literally everywhere. It's it's it's not it's not feasible. >> It's not going to work. It's not going to work. So, what you could do instead is have a bulldozer come in or let's say

36:32a really high wind and that's going to take down some of the trees. [snorts] >> Yeah. Okay. When you take down some of the trees, that's kind of like creating oxygen vacancies in something like a transition metal oxide like hnneium oxide or titanium dioxide, right? These are transition metal oxides that are metals bonded to oxygen. that's in the middle sandwiched in between the two electrodes. Right? If I've got an electric field, the oxygen which has these like it's it's just got these properties where it wants to go towards the negative electrode. Right? Um this is why oxygen is so reactive and why it rusts everything. Um the oxygen is going to start moving. When it starts moving,

37:13it's going to leave behind holes. >> Okay? So, it's kind of like the trees are coming down and now I've got a pathway. Now, my motorbike can move really fast. The motorbike in this analogy represents the electron. >> The electron is piggybacking on these vacancies. >> And and it it's and part of your point is being until the wind or the bulldozer comes and knocks the trees over, which is the oxygen leaving. >> Um [clears throat] the electrons can't move through. >> Yeah. So that would be a high resistance state, right? That's one of the states that we were thinking a memor should have. It's got a memory, right? There was no wind and so I am at a high resistance. Now all of a sudden I

37:54applied an electric field. I got a low resistance. Can I go backwards? Yeah. Just apply the electric field in the other direction. The trees stand up. Low resistance. >> Got it. Got it. >> Right. This is the mechanism behind the switching behavior. The set and the reset that we think of in memoristers. That's the memorister jargon. It's like you set it to the low resistance state. So then there's like conduction happening. So that's your one let's say. And then you reset it back to the high resistance state. So then the the trees are all up, the vacancies are back, the oxygen is back. I mean, there's no more vacancies, and the electrons don't have a highway to go through. >> And what's interesting about this is you can go, you have the ability to return

38:35to to return back and forth in between each of these states. >> Exactly. So now you've got a cycling, >> right? >> And that's something that you need for nonvolatile memory, right? >> Okay. That's effectively what this um vacancy mechanism is. >> Mhm. >> Got it. Now, a lot of times now we're going to try and use half oxide. The advantage here is that halfneium oxide is something that silicon chip makers have already been using in transistor gate stacks. It will it's already replaced um the silicon dioxide

Why hafnium oxide matters

39:07transition metal oxide in the famous high K dialectric constant metal gate transition that happened in like 2007. Um I wasn't aware of this obviously cuz I was in school but in 2007 like there was sort of a plateau on Moore's law because we couldn't get stuff smaller and then and then somebody decided hey actually we can use half oxide that reduces power leakage and it allows these transistors it allows these transistors to shrink even further without having a significant decrease in performance. The smaller you get, right, the harder it is to keep the same level of performance at larger >> scale. Yeah. Because you're trying to dissipate heat. The smaller you get now that the heat is starting to mess with

39:47the components themselves. And so this halfium oxide, it's CMOS compatible. It scales to very small dimensions. It switches fast and it shows really good endurance. Right? And by the mid2010s, all the major semiconductor companies, that's like Samsung, Micron, TSMC, they all have active memor development programs. and they still do. >> And and just to make just to make a quick clarification point, so chip makers were already using this for other chips prior to using it in memoristers. Yeah. >> And so part of it was like >> they had the industrial scale working with this particular material. And so

40:28they were like it would be ideal if we could use this same thing we already do at scale. >> Yeah. >> In another chip type. Is that is that is that fair? >> That's exactly right. And so the original paper that was by um Stanley Williams, that one used titanium dioxide as the transition metal oxide, but people have rapidly switched now to halfneium oxide as the sort of sandwich, the the thing in between the sandwich. Okay. Um now we've got halfium oxide. Okay, that's the meat in between the sandwich. What do we want to use for the electrodes? Usually we use tungsten and platinum.

Why conventional memristors fail when hot

41:02Why tungsten? Well, tungsten is has a very high affinity for oxygen. So, that's good. It's trying to get the oxygen. Um, and it's also really good at withstanding high temperatures. That's why the Thomas Edison light bulbs use a tungsten filament, right? And here we've got a zoom in of one of those filaments. It's made out of tungsten because it's not going to melt. Okay, so that makes sense. The basic idea now to make this memor is the following. We've got a top electrode. We've got a bottom electrode, tungsten, halfneeium oxide, something else. >> Mhm. >> And we use the oxygen vacancies in the half oxide to

41:45transport electrons and turn that highway on or off. >> Okay, >> that's effectively the idea. And I'm guessing the idea then here is whatever is on the opposite side of the tungsten depending on what it is changes the dynamics of the back and forth and so there's fertile ground to optimize for different use cases. >> That is exactly right and and let's let's stay here. Okay. So you you called it the top is tungsten the meat in the middle is hnneium oxide. What should go on the bottom? Okay. For a long time, people have been talking about platinum. Usually, you want some kind of inert

42:26metal. It's a metal that doesn't like react with stuff, right? Because obviously, if it reacts, that's going to cause performance degradation. So, platinum is a pretty good metal. This is where the veilance charge mechanism takes over, right? The problem is when you pump up the temperature at high temperature, the tungsten starts diffusing, >> right? If you get into a high enough temperature, the tungsten now starts jiggling around and the tungsten starts exploring the forest. >> Okay? >> And we don't want that. We don't want that. We just want oxygen. >> We we want only the oxygen to move around, create these vacancies, and then

43:06there's a highway. But if the tungsten starts moving around, all of a sudden, maybe you've got filaments of tungsten and maybe filaments of platinum that are coming together. And if those two meet, you've got a shorted circuit. >> Mhm. >> Now, there's no way to I mean, even if you cut down all the trees, you've just paved a highway >> for that. Yeah. Yeah. Yeah. And there's no it can no longer cycle. >> Yeah. Yeah. Why would anyone take the trees? You just go through the highway. Right. >> Right. You go through the tungsten and the platinum filaments. And that's been the big problem, right? They mix in ways that destroy this precise nanocale structure. Platinum and tungsten is not the way. Okay? You're going to effectively short circuit at high

43:46temperatures. At low temperatures, this thing is working >> great. Right? But you you start pumping this thing up to 500, 600, 700° C. All of a sudden, the the metal atoms in the two electrodes, they start getting agitated. This this is giving me uh a feeling similar to the hypersonics episodes we talked about where Navier Stokes gets weird >> at higher temperatures. And so there's different regimes depending on the temperature band you're working in and I'm I'm just making making the connection between the two things. >> Yeah. Different physics starts mattering. Right. Right. It's like it's like the the the titanium to titanium bond

44:27>> is now smaller than the thermal noise. Yeah. >> And so the titanium's like, "All right, I'm going to dip out. Let's start exploring. I don't know. I'm going to go this way." [laughter] And the platinum's like "Hey hey >> yeah." And and the structure, the way you've kind of laid it out now that that makes sense in terms of, "Okay, we we've discovered the fourth side, which is the memeister, we've figured out ways in which because of the way the industry was at the time, what kind of works and is good for most of the use cases we thought we needed. >> But then as we got to higher temperatures, physics gets different. And it does our the current regime of tungsten and platinum >> doesn't work. >> Doesn't work. >> Yes. Exactly. And now we get to the

45:08paper. Okay. This was out in science. Um shout out. >> Um high temperature meristers enabled by

The graphene solution

45:18interfacial engineering. One of the co-authors actually is Stanley Lee Williams who's the original memorist guy for the nature for four for four. Oh my god, what a legend. >> I thought that was a nice like like uh throughput. >> What a legend. >> Um so Joshua Yang who's at USC, he worked with Stanley at HP Labs and then now he's a professor at of electrical engineering at USC. So, this is a collaboration between, as you said, USC Air Force Research and Kumamoto. But it's like the same lab that created the Meisters is now like just upping the game. >> We're We're the best. What do you want us to do? >> Which is kind of cool. Their big idea is to replace that bottom inert electrode.

46:01>> Instead of platinum, we're going to use graphine. >> Okay. So, this is the one that we need to be more stable. Doesn't want to do stuff. >> Yeah. I want it to not talk to the platinum or sorry, I want it to not talk to the titanium. I want them to not be friends. Yes, platinum and titanium, they're friends at high temperature. >> Turns out graphine, even at high temperature, it's fine on its own. >> I'm good. >> Yeah, it's good. And for for very good reasons. So, graphine, it won the Nobel Prize in 2010. Um, it was isolated in 2004 by Andre Gim and Constantine Novyo Salv. No Salv. Okay. They were um in the UK. It's a very funny

46:44story where they actually they they took graphite, which is the stuff in pencil lead. They put it on pieces of paper and then they got scotch tape, put it on the piece of paper and removed it. And the physics between the scotch tape and the graphite that's on the piece of paper would create single layers of graphite, >> right? Single layers of carbon hexagonal shells. Graphite is a bunch of layers stacked together. >> Graphine is a single layer. Okay, it's it's an incredible story because it was scotch tape and a pencil and it won them the Nobel Prize. Sometimes curiosity and just trying stuff is the

47:25key to breakthroughs that will last a lifetime as we again have discussed multiple times on the show. >> And I mean since then we've done loads of research on how to make graphine. Obviously like if you're trying to do industrial scale, you're not going to be doing scotch tape and pencil, right? >> But you wouldn't have known. >> But you wouldn't have known, right? And now people do chemical vapor deposition to make graphine. Graphine has extraordinary properties. Okay? It's the strongest material ever measured. [snorts] It's an excellent electrical conductor. It's transparent. It's chemically inert. It doesn't really mess around with other things. Um, and that chemical inertness

48:06is the key property here. Side note, um, graphine for the longest time was overhyped. >> It was in Batman's suit, isn't it? Doesn't that what he uses in his suit? No, it's in it's in all the movies. >> All the movies. >> And then when when you go out, you're like, "Where is the graphic?" >> Where's the graph? [laughter] >> Okay. It won the Nobel Prize and everything, but like I think if you talk to a lot of material scientists, they're going to be like, "This thing was super overhyped because of those the it was the strongest material. It was a conductor. It was transparent. It was inert. It didn't really find a lot of chem like a lot of industrial applications. So, it's really nice to

48:48see that it's actually doing something. I'm sure there's going to be people in the comments that tell me that there's all sorts of other stuff the graphine does. It's just I haven't heard it, right? And >> I've heard I've heard of silken chips. I've heard of all these other things. >> Carbon nanop fibers like in F1 they use like carbon nanopers in the but that's not graphine >> and carbon nanopers never won a Nobel Prize. So, I'm just saying >> bad PR, >> right? Um this this is a totally new purpose though. >> Mhm. Because a lot of times people like to use graphine for as you said the the myar and Batman suit like because it's strong or it's like electrically conducting. Here we're exploiting the chemical inertness of graphine and the

49:29and it may be finally getting its moment in in the in the light because the the implications of this uh could be just so enormous.

Testing the device at 700°C

49:41>> Yeah, it really could. It really could. So let's now get into the paper. Let's talk about figure one. The graphine solution. Uh figure 1 a on the top left. That is their memor. Okay. You've got a little tungsten electrode up top. You've got halfname oxide in the middle and you've got a graphine electrode on the bottom. Okay. And that's effectively what's happening. The tungsten doesn't want to mess with the graphine. We're going to get into some of the cool physics about that later. on the right hand side is what I want to show. You see that curve that hysteresus curve that's characteristic of your >> memeister. So that is where they're showing the memeister.

50:24>> All of the colors are from from 0° C all the way down to 700°. >> And notice the curve maintains its shape. >> Yes. >> All the way down to 700°. This is crazy. >> Yeah. Yeah. Yeah. Yeah. And because we're talking about 2 and a half 3x >> Yeah. >> the current level that we can >> Yeah. And if you I mean usually in in temperature we always as physicists like to talk in Kelvin, right? 0° C is going to be like 270 let's say 300 300 Kelvin. 700° C is 1,000 Kelvin. So this is three times the thermal energy that this thing is able to withstand because Kelvin is

51:04really how you scale like how much kinetic energy is in the system, right? like um at at 1,000 Kelvin or at 900 Kelvin, let's say, your atoms are moving with three times the kinetic energy. >> Yeah. >> As the atoms at 300 Kelvin, >> okay, at room temperature. So this is this is crazy. >> Yeah, it's it's Yeah. >> Yeah. On the on the bottom left plot, you're seeing the time axis on the x- axis, right? >> Yep. >> That's 10 the 4 seconds. You're maintaining those two states of high resistance, low resistance. >> So the cycling is maintaining.

51:44>> Yeah. >> That we talked about. The forests is going up and down. >> Yeah. And it's fine. >> And it's no problem. >> And it's not even coming together. >> Yeah. >> It's just flat, >> which is Right. Right. Right. Like there's not a de like there there's not a the collapse the it's not approaching the the circuit the short circuit. >> No, it's not. Yeah. Short circuit would exactly mean that like they're both coming to the same thing. We're not seeing that at all. at 700°C. This is cycling at 700°C. >> So we have it's stable. >> Yes. And on the right hand side we've got endurance of cycles. The x-axis there on the bottom right that's in log scale. >> Holy. >> Okay. That's not a linear scale. Each of those tick marks is 10 endurance cycles, then 100, then a,000, then 10,000,

52:26100,000. It goes all the way up to a billion cycles >> before we start to see. >> Yeah. And even then that could just be noise because there's there's there's enough like you know jitter in the beginning. >> Yeah. Yeah. Yeah. Yeah. It's not clearly this is incredible, >> right? Because what we're sort of saying there's two things here like one stability at high temperature which is C and then D is stability over time which goes back to the the two things. And I'm kind of getting what you were saying before about why the fundamentals of the structure impact both of these variables. >> Exactly. It's impacting both the temperature side of things which is in

53:07the B and the time side of things because it's like stable. The whole thing is stable. And actually the next one which is in I think this is figure two in the paper. This shows a schematic. You always want a cartoon for all us stupid people like me who are trying to understand what's going on. On the left is a old memeister. Okay. Tungsten, halfium oxide, platinum. At high temperatures, the tungsten is migrating and creating a paved highway. >> Mhm. >> So the electrons will just move through that. >> Yes. >> Cuz that's the path of least resistance. >> Yeah. No, no one's going to take the vacancies, the oxygen vacancies. Why would you do that when there's a highway right next door? M um it's like the Cars movie, you know, when the interstate was

53:47formed, [laughter] nobody went through Route 66 and then your town got wrecked, right? Okay. So, on the right hand side now, um there's no highway because the tungsten is trying to diffuse, but it just bounces back from the graphine. The graphine's like, I'm good. >> I'm all carbon >> and I'm good. >> It's It's like uh California's uh high-speed train. Yeah. >> 86 miles of track, but you can't really use it. >> Yeah, you can't. Yeah. I know. Where is it going? Where is it going? Nowhere. >> Nowhere. Right. And so and so that's exactly right. Yeah. And so the the the tungsten is sort of just like trying. It's still diffusing around. The atoms are still diffusing because of the high temperature, but it can't bind to the

54:28graphine. And so the only way through is with the vacancy centers that are created through that >> hiker's paradise. And that's where the electrons shuttle. And so now we have exquisite control of the resistance state. the the the one on the right here, the the the hiker's path that you talk about that structure which is the forest, the mo the bike analogy where we have the ability to control when they're up or down. That that that's the point is that that's the le the lever that gives us uh programmatic control for

Proving what happens at the atomic interface

55:00lack of a better term. >> Exactly. Exactly. And Professor Joshua Wang, who was the principal investigator, he described this interface interaction between the tungsten and the graphine as oil and water. It's just like They're just not going to mix. Right. Makes sense. >> This is so good. >> It's it's it's pretty cool, right? So, now let's um you know, to get into a paper in science, >> you can't just have a cartoon. >> Yeah. [laughter] >> Right. That that's for me. >> But but uh in order to p in order to publish in science and convince your competitors that this is legit and all that other kind of stuff, you need actual analytical verification. So, let's go to the first analytical verification. This is high resolution trans transmission electron microscopy,

55:42electron microscopes. We've covered this a lot. Instead of light, we're using high energy electrons to really get down to um you know tiny tiny nanometer level resolution. On the top is a traditional memeister. Okay. You can see after after like not a lot of time the curves collapse. Yeah. >> Right. after just like 2,000 seconds, which is like less than an hour. Um, there's no memor. >> It's the slope of uh a mammoth. >> Yeah. Yeah. Yeah. Exactly. >> Straight down. >> Yeah. Yeah. Chair 36, [laughter] I believe. Correct me in the comments. I never go on those chairs. So, that'll be you lunatics. But on the right hand side

56:24is the transmission electron microscopes. And you can see at high temperature, you see this like mushroom cloud that's forming on the right hand side. >> That's the platinum and the tungsten meeting up. M and so instead of sort of maintaining the same like substrate texture or orientation this there's just this disturbance in the force. >> Exactly. There's visually see it. >> Yeah. And you can see the halfnam oxide in the middle right the meat but like the platinum is just like going through. Why would you why would you go through the half oxide right on the bottom >> is our new meister with graphine. And there you can see even at high temperature, even after all of that cycling, >> the graphine layer >> the graphine layer is inert. The

57:06tungsten layer is inert. Nothing is happening. This is incredible. There's no chemistry happening. >> And it's it's also so verifiable. >> Yeah. You can literally look at it under the microscope and I can see >> and you can see that it's happening. >> That stupid analogy is like imagine you have a BLT sandwich from a deli and they put the the thumb the the the toothpick through it. In the first case, it's like, you know, the toothpick through where you can see it disturbs the nice layers you have of your tomato, bacon, lettuce. Yeah. >> And it looks like a very nice undisturbed. But but the point here is the the high resolution imagery that is able to validate that this graphine based meister actually sustains at high

57:48temperature is undeniable. >> Yeah. Yeah. Yeah. It's like now you don't need a cartoon. This is the thing that the cartoon is literally what it is. Right. right? Like you're just seeing it. I can see it. Um here's another way to see it. You can use energy dispersive X-ray spectroscopy. The idea here is that um like so during in the previous photo we we had transmission electron microscopy. So you're like sending these high energy electron beams, right? That strike the sample. Now when those electrons strike the sample, it can transfer energy and violently kick electrons out. And when the electrons come back, they're going to release X-rays, right? And if I monitor the X-rays during that electron

58:30microscopy because every atom is unique, the X-ray signature is going to be unique. Like in the transmission electron microscopy, people could be like, oh, like there's some weird thing happening. How do you know that's tungsten? How do you know that's plat platinum? Right? It's just a gray image. Well, now we can tell that it is in fact tungsten that is leaking through because only a certain energy comes from tungsten. And we're seeing that energy spike. >> Yeah. >> In the old memorister, but in the new memor there's no tungsten spike. >> Tungsten spike. >> The W, by the way, for those watching, W is the chemical symbol for tungsten >> because of Latin. And it so it's

59:12interesting because I I think part of this is like you you're doing both of these things almost simultaneously because because the the electron microscopy is triggering an event that you track with the X-ray spectroscopy. And so you're getting both the visual and the the signature. >> Yeah. Exactly. >> That allows you to identify the uh the composition of what you're looking at and can see visually. >> Exactly. And so if you look in the platinum electrode, there is tungsten there. If you look in the graphine, there's no tungsten, >> right? >> Undeniable. >> Um, there were also some calculations that were done from first principles. >> So, um, you've got electron densities with density functional theory. Density functional theory is something that we

59:52covered in our America 250 episode. It won the Nobel Prize. This is what I mean by density functional theory. Okay? It's like you don't have to worry about every single electron. You can just be like the electron clouds are like a potential. What happens when a tungsten atom goes through? Does it bind or not? On the right hand side is what platinum looks like. These giant sort of honeycomb structures where the dorbital, which is the stuff that um these are these large dorbitals of metals, right? They're extremely large. The tungsten is fine kind of just binding to one of these dorbitals because they're so they're so big, right? And there's a lot

1:00:35of vacant spots where the tungsten can go and bind. On the left hand side is graphine. Pristine hexagonal structure. Um I think it's only just like sp2 binding or maybe sp3. I I forget my chemistry. But in any case, the the bond lengths in between the carbons are smaller. The graphine electron clouds are a lot smaller. So the tungsten has nowhere to go. >> It's incredible. >> Right? There's the You see the fat little like pimple? Yes. >> That's the tungsten atom and it's trying to find somewhere to bind >> and it can't. >> Yeah. It's like the grill on like your microphone or a speaker cover. Yeah.

1:01:15>> Uh versus like you described like >> a bunch of donuts that got blend or like you know Hawaiian rolls that got that got baked together with no clear actual like structural like you know like a well-defined small structural positioning. Exactly. Um, and it's part of your point that the the the smaller size of the electron cloud is a material aspect of why >> the because the tungsten is just not small enough to >> Yeah. Yeah. The tungsten is not small enough. The smaller size also means that you need a higher energy to get in there. >> That makes sense, right? And um and and all of the stuff is is taken. >> Yes. >> Right. With with graphine because of that hexagonal structure, it's just carbon on carbon on carbon. All of the

1:01:56carbon atoms are happy. They don't want anything else. >> They don't want anything. >> Right. Yeah, it all. Oh, that's so good. >> So, it's all coming together. And the final thing that I want to talk about is um the nudged elastic band. >> Uh this isn't the final thing, but I just go go I'm going through their figures because they're just so cool. Okay, so [clears throat] on the left hand side, again, that's the old Memorister. Um, this is an activation energy diagram that shows if I want the tungsten to bind to my platinum, I need to [snorts] go over this hill and then and then there's like a valley on the other side where I'm bound to the platinum. The height of the hill is.3 electron volts. >> Mhm. [clears throat] >> And then the the there's a there's a

1:02:37nice valley where I can just like sit. On the right hand side is the same diagram but for graphine. The height of the hill is now 1 to two electron volts. And there's no nice valley, right? Like if I'm up there and I'm jiggling around, I'll just come right back. >> Yeah. >> To my original state. >> So again, same idea. It's really hard to bind to graphine. >> It's really good. >> Okay. >> Now, one thing that I was thinking about was there's kind of a paradox in my head, right? Like the high resistance state makes sense. high resistance meaning um the there's no oxygen vacancy highway

Why both memory states remain stable

1:03:16and so um like my electrons aren't going through but everything is kind of inert right now let's say I have oxygen vacancy highway in that low resistance state what prevents the oxygen vacancy highway from moving around >> at high temperature right at high temperature if I've got these like vacancy lines where the electrons are going through, but those vacancies can move as >> we talked about earlier. They're not stationary. They're not fixed or >> Yeah. Yeah. Yeah. So, what's preventing those guys from moving around at high temperature when there's all this jiggle? It has to do with something called phase separation. There was a paper about this where the system

1:03:58reaches thermodynamic flux equilibrium where there's a net migration of zero meaning that the material separates into an oxygenrich insulating phase and then an oxygen poor conducting phase. And it's that same argument about oil and water. There's going to be parts of the hneium oxide that have a lot of these oxygen vacancies and then there's going to be parts that don't. And those guys don't want to mix together >> because they become like oil and water. Now, how can we tell in our new memory? Well, they did electron energy loss spectroscopy, which is similar to the X-ray stuff that I was talking about. Like the electrons are now coming out, and then we measure the the energy of

1:04:40the electrons. The main thing that I want you to see is in the low resistance state, which is on the bottom. So, the color tells you how much oxygen vacancy there is. The darker it is, the more oxygen vacancy there is. So that's why the low resistance state at the bottom that has it's kind of darker because there's a lot more oxygen vacancies. The the key thing that I want you to notice is on the bottom we're plotting the number of oxygen vacancies as a function of space. Like we're going we're moving along in nanometer position. We're moving along our meister and we're plotting how much oxygen vacancy there is. If you if you expect a lot of

1:05:20mixing, it should be at equilibrium. But if there's not a lot of mixing, there should be clumps where there's a lot of oxygen vacancy and then not a lot of oxygen vacancy >> because of this oil and wild oil and water you just talked about. >> And that's what we're seeing. We're seeing a very bumpy profile, right? There's certain spots where there's a lot of oxygen vacancy and then other spots that there aren't. If this stuff mixed, it would look like the middle plot >> in that high resistance state. >> Mhm. >> Right. >> Mhm. Yes. >> But because it's not mixing, you get this valley and mountain landscape. And so part of what we're saying is there's two dynamics here. There's both a dynamic between the inter how the tungsten and the graphine relate

1:06:00>> and then the the the halified oxygen. >> Yeah. Um, halfneium halfneium oxide should be halfneium oxide in the middle also has dynamics that not that are com that combine with the relationship between the tungsten and the graphine such that the graphine doesn't want to create the highway and the oxide in the middle also doesn't allow movement. Yeah. And so both of these things combined are are meaningful for the impact in terms of from an industrial or performance perspective. And it's both things not just the ends the the ends of the sandwich. >> Exactly. Yeah. So both the high

1:06:42resistance state which is the one the or sorry no the high resistance is the zero cuz that's like the the electrons aren't moving and then the low resistance state where there is current that's the one. Both of those are now stable. >> Right. Right. >> Uh for for dynamics that we just walked through. >> Mhm. which is which is which is an important point because I was gonna end up asking but what about the middle stuff? >> Yeah. Yeah. The middle stuff is also oil and water. >> Right. Right. Right. >> It's kind of cool. >> Um and they did a lot of benchmarking of like okay what what is all the state-of-the-art and then where is this >> and um it's kind of crazy retention which is how how long it takes and the temperature. That's part A. You want it

1:07:22to be high temperature high retention. So you've got a desired corner >> that's the shaded and there's a [clears throat] star where the current work is and all of the other riff raff >> is not even close >> is not even close. Um you've got the writing endurance in B versus temperature again the only star in the desired corner is this work. >> Um C the onoff ratio versus the temperature desired corner is us. the spiderweb diagram. Um the yellow is the current work and it's always >> larger >> in temperature, in retention, in endurance, in the device size, all sorts of stuff. Now, there is a little bit of

1:08:04uh photoshoppery happening here, [laughter] right? Because I could easily make the desired corner like elsewhere, but but in any case, it's in by every metric, it's always in it's the most in that corner. So whatever desired corner you made, it would always be the most >> out there. Right. >> It's the closest to the optimal regardless of made whether you made the desired outcome smaller or large or whatever. >> Yeah. Yeah. Yeah. But a game recognizes game cuz like I've done that. >> I know what you're Yeah. I'm going to make this square right here so that all the other >> But like mine just mix it [laughter] in. >> Yeah. But in this case, I think it's like fairly obvious and I don't think they're doing any shenanigans. I just

1:08:44thought that was kind of funny as a plot. Um, and finally like to cuz I think they knew that people would call BS right? >> So they took a video of their lab. >> They're like proof. >> They're like this is proof. And the and the thingy shows, oh, 700 degrees Celsius. And they've got their they've got like a computer with their with the two electrodes on the on the two on the the the titanium and the graphine. And they're measuring the onoff ratio and they're measuring the hysteresus curve. And here you can see the hysteresus curve. Like, okay, I turned the voltage up, I turn the voltage down. I'm getting this hysteresus. Guys, this is happening in

1:09:25real time. I'm not making this stuff up. Like I think it's really cool that like this is part of the supplement on the science website because like um whenever you publish right there's the main figures and the main text and then you provide supplementary data and supplementary figures. A lot of the figures that we've seen here um in today's episode are from the supplement >> and this is a supplementary video that they show. Okay guys, here's my hysteresus curve. It's definitely a memeister. This is happening at 700 degrees Celsius and it's definitely working. >> Pix or it didn't happen. >> Yeah. >> Uh in uh front tier science research this but

1:10:05it's also kind of like I I think one people one thing you know as we always talk about is scientists are some of the most skeptical people especially when it's other people that say they did something. >> It's like why didn't I think of that? >> I [laughter] don't I don't believe you. >> Yeah. >> But here Yeah. It's like, okay, here's 704° C, guys. There's my thermometer. [laughter] >> God, this is so good because, you know, and I think you've you've done a good job of how walking us through to kind of understand that there's a fundamental insight that happens that then you you take to the end of possibility or the end of kind of the story in order to get to something that's practical. And it

1:10:48was a very it was maybe subtle um for folks who are material scientists and who work in the space but if you're not in this lane um it can seem like oh this is a subtle point that has such huge implications and >> they've used graphine as the other part of this layer here. Are there other things that have different >> performance characteristics? Yeah, maybe we can now start maybe we can now start um replacing the tungsten. I don't know, >> right? And then and then what does that do? >> Yeah, it just it opens up all sorts of possibilities,

The future of high-temperature computing

1:11:25>> right? >> So now let's talk about the future. Okay, what would this current research have to do with how we move forward in computation as a society? We've talked about the vonoman bottleneck for a long time. This is the tyranny of the processor having to be separate from our memory. Modern architectures waste a lot of power and time trying to move data between our CPU which is where all of the computation is happening and the RAM which is where the data and the memory is stored. This was vonoman came up with this and it's the bedrock of almost all of the computation that we have in our society today. Now, with

1:12:06neural networks, this is especially bad because you've got to load the weights and then you've got to do the computation, then you've got to load a new set of weights. So, you got to dump the old weights, right? And all of this takes a lot of time. It takes a lot of energy, dissipates heat, and what exactly is the computation that happens in these AI systems? When you've got a neural network, what we're really doing fundamentally is we're relying on matrix times vector multiplication. The vector is your state and the matrix is the weights of your network. You multiply the matrix. You multiply the matrix to the vector and then you get a new vector

1:12:46that tells you how the state is changing. At the end of the whole thing, you get a giant vector. In in in the example of of LLMs, you have a giant vector with all of the next possible tokens and you pick usually the one with the highest probability. Okay, how does matrix multiplication work? How do you multiply a matrix by a vector? For those from linear algebra, you already know this, but let's just do a little bit of a review. Okay, let's say I've got a 3x3 matrix on the left there. I've got 112 213142 and I want to multiply it by a vector 312. 312 is the vector that's my state of like next tokens, let's say. And the matrix tells us how to change that

1:13:29vector to create a new probability vector with the new tokens. Okay, what you do is you go row by row and you take each row, you multiply it by the vector. So the one gets multiplied by the three, the one gets multiplied by the one, two gets multiplied by the two, and you add it all up. That's your first row. Then you go to the second row, add it all up. Third row, multiply, add it all up, and you get your new vector. Okay, that's fundamentally what most of the AI computation that we all know and love. I guess that's what that is. Okay, that's what's happening. >> Keyword, I guess. >> Yeah, that's what's happening. Now,

In-memory computing and AI

1:14:10we'd like to do better. We'd like to not load the weights every single time in order to do this computation. >> And this is part of the reason why, for example, the cloud versus local with AI models and stuff. It's why if you want to do it locally, you need a very powerful machine because it has to do this. >> Yeah. It has to store all of those matrices and it's got to do all of that computation. That's why GPUs are so important because GPUs are just insanely good at matrix multiplication >> at that specific >> at that very specific task. Okay, they're not good at like doing normal emails. But it turns out Nvidia made the big

1:14:51gamble way back in the day that actually graphics, which is matrix multiplication, is going to have a use for later. And it turns out, yeah, >> in things other than video games, for those of who are gamers, we've been very familiar with GPUs for a while, and now everyone is making custom PC builds unnecessarily expensive. >> Yeah. Yeah. Exactly. So, instead of doing all this vonoman nonsense in instead we can do something called in-memory computing. In-memory computing meaning I don't have to move anything. Okay? we can use something like a resistor to do the computation without changing

1:15:31anything. So again, let's talk about how that would work. I told you about how matrix multiplication works, right? Like matrices, you multiply it by a vector and then you get a new vector. Well, now what if we had the old vector come in like wires? Okay, and this is what we see at the on the on the right. And let me let me see if I can say this correctly. Okay, so I've got a matrix and I've got an old vector and I want to make a new vector. And the way that I do that is the rows get multiplied to each of them and then I get added up, right? I can use the rules of electronics V= IR and the summing of voltages and the summing of currents to then just do that using physics. Here's what I do. My old

1:16:14vector is represented by the old by currents that are going at the very top. The old voltages, I should say. They're they're each of the wires at the very top is held at a certain voltage. Right? I've got resistors that are attached to these voltages. And then I've got a new set of wires at the bottom. Because of electronics, the new current, it's a new set of currents at the bottom. Because of electronics, right, the new currents are going to be related to the voltage at the top and the resistance in

1:16:54between. And all I have to do is add up the currents. >> I'm so mad. >> Right? >> In this case, it's not really resistance. It's one over resistance. It's the it's the conductance. And the and then the current is just voltage multiplied by conductance. But it's the same thing. It's just V equ= Iir the other way around. And all we're doing is using the physics in place, right? We don't have to move weights around. As long as the weights are the same, I can use the physics to be like voltage multiplied by conductance gives me a current. Voltage multiplied by another conductance get gives me another current. And so my new vector is going to be the sum of all of those currents. And there you go. Matrix multiplication

1:17:34using physics. So I'm just I'm so mad. That's so that's that's so good. And the part of the point here is then then you can get around the BOMAN bottleneck as a result of that because you're not having to transition between your compute layer and your memory layer. >> Yeah. And all you have to do is the weights have to be programmed into the conductance. The input is the voltage and the resulting multiplication comes out as current. I'm just using Ohm's law and Kirkoff's junction rules. The stuff that you learn in like intro physics right now. Are we capable of doing something like this? Well, this particular paper itself made a 32x32 analog resistance state and it works at 1300° F. It works at low voltage. 0.5

1:18:18volts is all you need. um 30 nanocond switching in the speed and it definitively it it it proves that you can be fully capable of storing complex multi-level analog neural weights for whatever AI interface that you're working with. It can do it in extreme environments 1300° Fahrenheit and the standard digital GPU which relies on all these transistors and the cos and all that other kind of stuff the the traditional GPU is going to get fried. But if I use this memory where the the conductances, the weights are stored using this memory, now all of a sudden I've got something that can actually work. Golly, it's so it's so

1:19:02good. And it it's this interesting combination of both things which is the um fundamental discovery around the

TetraMem and commercialization

1:19:12memorisser construction for high temperature plus the concept of uh in-memory computing >> together. Yeah. >> Um >> there's a there because it's it's not just one or the other. Both things give you now the single chip that can work in high temperature that can you know run comparative uh comparably for a long time >> faster >> faster um and can do both your compute and memory at the same time. >> Exactly. And of course the authors have made a startup. >> Okay. Of course there's a startup called Tetrame. It was co-founded by Joshua

1:19:52Yang who's the PI of the paper that we're talking about along with um Chiang Fe Xia and the other authors. Um and it's already working to commercialize this crossbar memorist device to replace power hungry GPUs, right? >> They're getting acquired immediately. >> Yeah. No, especially with this >> immediately. >> Now, let's talk about what this has to do with data centers in space. Okay. >> Okay. Elon wants to put data centers in space. Why? because it's probably possible and I mean it helps SpaceX obviously and it's not just SpaceX that's trying to do this right Google is trying to do this there's a bunch of startups that are trying to do this and

1:20:32it's looked impossible for a very very long time because of a sort of fundamental physics quandry is because these GPUs are highly reliant on cooling mechanisms because the GPUs can't withstand and a lot of heat. That's why whenever you hear about like all these data centers taking up um local water, the water is being used to cool down the GPUs effectively. At the end of the day, that's most of the water being used is just to extract heat from these GPUs and put it out into the environment. That's why the data centers, the climate around the data centers, there's like a microclimate that's created and the

1:21:14temperature gets raised and then all the environmental people are like, "Okay, what's happening in the flora and fauna?" which is completely reasonable, right? I don't want to hurt like the planet and especially the planet that's right around those data centers. So, >> let's talk about how treat heat transfer happens. There's three types of heat transfer. Okay. There's convection, conduction, and radiation. Convection is hot stuff moves around to the cold part. >> Mhm. >> And then the cold part goes down to where the hot hot stuff is. That's how like if you put a pot on a stove, the the water gets heated uniformly because the hot packets are moving and the cold packets are moving. So that's convection. Conduction is [snorts] the

1:21:56atoms themselves jiggling around to conduct heat from one spot to the other. That's why the handle gets hot. And then finally, the most inefficient part is radiation, which is literally photons coming out. M like the reason why a fire is hot, but if you put your hand in front of the fire, all of a sudden you don't feel that hot is because you've stopped photons from getting to your face. Okay? And that's why like when when a shade is so important because now the photons, the infrared photons from the sun are not hitting you and then all of a sudden you feel like 10 or 20° cooler. >> You're right next to the sun, right? The shade is only shading your part. Well,

1:22:37the reason why it's so effective is because the photons aren't getting to you. The convection and the conduction still is, right? The the hot pockets of air are still moving towards you, but the photons are not. Okay? So, that those are the three types of heat transfer. Now, terrestrial data centers, they mostly use the first two. Mostly, it's convection. That's why you've got water pipes going through because the water gets hot and then you move the water to a colder spot. It dumps the heat and you cycle it back. back and so on and so forth, right? Um, terrestrial

Why put data centers in space?

1:23:10data centers have issues, right? Where is the power going to come from? Where's the water going to come from? Who's going to build it? >> Yeah. Who's going to pay for it? >> Yeah. There's a lot of nimbiness going on. And in this case, the nimbiness is a good thing, right? We don't want >> I don't want data centers in my neighborhood. Okay. >> I might want the AI, but I really don't want the data [laughter] centers, right? And so that's why people are looking into space now. People have already started looking. This is StarCloud. It's a startup that's backed by Nvidia. Google is doing its own thing. And there's a lot of smart people betting on this thing. Here it shows a 5 gawatt data center. >> Crazy. >> But the radiator that you would need to dissipate all that heat with radiation

1:23:52is something like a 4 km square. >> And you have real um you know getting any payload into space is cost prohibitive. Yeah. >> 4 km wide will take a modular system, multiple launches. Yes. The cost of launches is going down. >> Yeah. >> Uh but >> still it's 4 km worth of >> stuff worth of [snorts] stuff, right? >> It's not a joke. >> It's not it's not trivial, right? And there's several important papers that are coming out trying to plan for this kind of thing like how do we develop carbon neutral data centers in space? I want to highlight this because this is in nature electronics. This is a top journal. Um, it's not impossible.

1:24:32There's a lot of rhetoric out there that's saying that data centers in space is impossible because of the vacuum of space and people need to take undergraduate physics. It's not true, right? Okay. It's hard, but that's a key distinction from it's impossible. Okay. We I mean, we obviously do it, right? The the International Space Station is already doing this. If we look at a photo of the International Space Station, the curved panels that we see, >> those are solar panels that are facing the sun. >> The the flat ones, those are radiators. Those are radiating out heat into space. So that the astron the astronauts that are on board don't get cooked alive. And all of the electronics that is, you

1:25:14know, doing computation, there's several computers up there, those don't get cooked alive. The way that works is using both convection and radiation. So you've got convection in that they use ammonia instead of water because ammonia has a low freezing point. If you [clears throat] use water then the water is just going to freeze when it gets to the radiator and then like what? So so so that's why you use ammonia. The ammonia uses convection to take the heat to the radiator and then the radiator takes that heat and radiates it out into space. Okay. Now what we want to do is make this thing cheaper and that is perhaps an engineering problem especially in the context of the paper that we've just covered. >> Mhm.

1:25:54>> I'd like to I'd like to pose that as a

The Stefan–Boltzmann law

1:25:57challenge and then see if we can work through with just some like first principles back of the envelope thinking. Okay. The technology poses three advantages. One, let's talk about the physics of thermal radiation. The reason why everyone's saying that it's going to be impossible, right? Thermal radiation is governed by something called the Stefon Boltzman law, which states that the energy that is radiated by a surface at a certain temperature is proportional to the 4th power of that temperature. This is why it's hard. Okay? It's not T^2, it's T 4th. So, if I double the temperature, I'm increasing the amount of power output I need or the the power output

1:26:38that's going going out. um by a factor of 16. >> Okay, now this is actually a good thing. Okay, in the context of this current paper, now standard silicon processors, they've got to be relatively cool to prevent failure, right? But you can only allow that chip with the standard silicon >> um to run at like tops 100° C. At 100° C, right, uh let's say 400 400 Kelvin is the tops that you can do. Um, at 400 Kelvin, the amount of radiation that I'm letting off is going to be some amount, some amount of photons are going off. Now, what if I

1:27:19could make that go up to 700° C, 1,000 Kelvin? So, I've effectively doubled the temperature that I'm working with. Well, the number of photons, the amount of power that I'm dissipating has now gone up by a 16, >> by a factor of 16. If the amount of power per unit area has gone up by a factor of 16, then in order to dissipate the same amount of power, I only need a 16th >> the size of radiator. >> Yeah. >> Yes. The size of the radiator goes down >> the higher the temperature that you can operate. And this is crucial for putting the stuff into space. >> Yeah.

1:27:59>> Right. Because of what you said, it's it's hard to put stuff in pa in space. >> Yes. >> It's hard. It's expensive. I'd like to put less stuff in space. This could be the key, right? Because now I can operate things at a much higher temperature, meaning that the amount of watts that are going out as radiation into space is higher, which means the amount of area that I need for my radiator is smaller. This is so interesting. And it is the it is enabled by this high temperature >> meister >> and the fundamentals that we just walked through that again now give you a different substrate a whole different computing paradigm when you think about the concept of a data center in space.

1:28:40>> Yeah, exactly. And I did some back of the envelope calculations, right? So let's let's say um I need a radiator for a 1 megawatt space server. Okay, that means 1 megawatt of electricity is coming in from my solar panels. Okay,

Calculating the size of a space radiator

1:28:56now at standard 30° C, so that's 300 Kelvin. If I want to dump 1 megawatt of heat, I would need 2,000 m squared. That's 8 tennis courts. And I want to dump just as much energy that's coming in, right? Because if if I don't dump the energy that's coming in, then my temperature increases. And then the electronics are not stable. My satellite keeps increasing in temperature. So whatever energy is coming in, I'd like to dump out. So 1 megawatt in 1 megawatt out the what matters is the temperature that I'm operating at. Okay. Now if 1 megawatt is coming in and now let's say I'm operating at 300° C, that's around 600 Kelvinish.

1:29:37My area has gone down to 150 m squared. I've gone from 2,000 m squared to 150 me squared. That's a reduction of 90%. Right? If I go even further because this thing was at 700° C, right? If I go even further to 600° C now, meaning 900 Kelvin, now my radiators only need to be 30 m squared. And that's totally doable. >> Yeah, that's very reasonable. >> Right. That's that's already in the ballpark of stuff that we've already put out there. >> Right. Right. Right. Right. That's so fat. It's it for me it's a little counterintuitive where it's at the higher you know when you before we started this the higher the temperature

1:30:18the less surface area you need to dissipate >> the heat. It doesn't feel >> Yeah. As long as the stuff is not cooking. >> Right. >> Right. But like it kind of makes sense because um like if you think about like you know those like electric stoves with the coils >> Yeah. >> right? The the higher the temperature the more the more heat is coming out. That's the whole point of how it cooks stuff. that. No, >> right. I mean, no, I guess that's not true because there's like contact, but you get what I'm saying, right? Like, if you're around that thing, >> Yeah, I know what you mean. >> It's like there's there's a lot more photons that are coming out. So, you're dumping a lot more energy and as long as you can live >> in that >> at 700° C, you're good to go. >> And it's just because we've never been able to think about it because we've

1:30:58never been able to operate that temperature level. So, it's never was even a possibility. >> There were no chip components that could withstand that temperature. So everyone was doing these calculations at the lower temperature being like oh I need tennis court sized 4 km squared size but now I've reduced stuff down to like 95 99% the area just because my components can withstand the higher temperature my components my computer can be hotter and so the amount of radiator I need is a lot lot lower by t to the 4th >> this is so I mean you know again there is the industrial scale and all of these other, you know, engineering. But, but to the point you brought up earlier, a

1:31:40lot of the narrative and rhetoric has been dismissive on its face in a way that doesn't even take into account

Are orbital data centers actually possible?

1:31:51what we're doing at the frontier. I mean, this paper came out in March. Yeah. By the way, and this discussion about data centers in space has really heated up as we've gotten into the summer. Yes. uh which was after this paper had come out and most of the folks who have been like it's impossible >> you know certainly one from a fundamental physics perspective >> it's not impossible >> it's not impossible and then also from where we are on the frontier both from a basic research and an engineering perspective it's a very different paradigm than you know just talking about putting the same thing we have in a data center on the ground >> and just putting it in space >> exactly and if you combine that with the fact that payloads to space are going to get cheaper as you the amount of payload that I need is

1:32:33going to go down. This is totally a doable thing, right? As long as we keep innovating and we keep going through with high temperature memoristers, high temperature memory, high temperature computation, that's going to be the key, I think, to unlocking data centers in space. >> Yeah. Yeah. Because you can have a lot more smaller, you can have, you know, much smaller space servers >> uh with high output capability. Um meaning we're not going to be, you know,

Power efficiency and radiation resilience

1:33:02looking up at the night sky and seeing, you know, >> Yeah. >> these giant floating radiators. >> Yeah. And and I I mean it it perhaps this is is going to be bad for astronomy, terrestrial astronomy, but there's there's got to be ways around it. I mean, it's already bad with the Starlink satellites that are up, but does that mean that like we don't provide internet to like these remote areas? Like, you know, technology and innovation comes at a cost, and we're just going to have to figure out ways to battle. >> And with the lower cost of payload, you could argue we could have a much larger array of space-based instrumentation. >> Exactly. >> That exist that negates the need to have terrestrial observatories. Again,

1:33:44there's some caveats there. >> There's a lot of caveats. A >> lot of caveats. Right. But that, you know, they're trade-offs and we'll have to continue to >> Yeah. Maybe we can make the orbits like avoid >> Chile, which is where like half of everything is, [laughter] right? >> If you just avoid like Hawaii and Chile, like half of the astronomers are going to be happy, right? >> You know, so um I mean that's going to be hard given that it's like high Earth orbit and whatever. But in any case, >> the the other thing that I want to bring up is this is not just an advantage in terms of cooling. Okay, there's two other things that this new memory technology would give us >> for data centers in space. One is the extreme power efficiency. I was talking

1:34:25about how it only takes like very little amount of voltage in order to do that computation with a memory. Now satellites operate on extremely strict power budgets. You can imagine right often it's under 20 watts and in standard AI chips nearly every single watt of electricity is consumed directly into heat but memoristers use this in-memory analog computing they don't have to do this von bottleneck right so if we're applying memoristers here then that complex mathematics instantaneously happens without shuttling the data back and forth >> now you're consuming a fraction of the power of a traditional GPU and that's going to bring down the amount of power

1:35:06that you need to dissipate again. >> So, this is both you can have the radiators be smaller and you can have the solar panels be smaller. >> Exactly. >> Because you don't need as much power intake. >> Exactly. Or with the same solar panels, you can do a lot more. Right. >> You know, either way, right? Efficiency is always great. >> Yeah. Yeah. >> Both in power dissipation and in terms of computing. And finally, the other big thing that people talk about with data centers in space is like normal GPUs, a cosmic ray comes in. >> Yeah. I don't know where that what if that bit was a zero or a one, right? Like what? So like a a galaxy decided to fart and then and then now like my AI query is giving me nonsense. That's not

1:35:47great, right? You need electronics that is able to sustain performance in a high radiation environment. >> Memoristers are great for this because memorist store memory by physically moving atoms and creating structural bridges, right? this like veilance charge mechanism rather than relying on these easily disrupted pockets of electrical charge. So now you can imagine a cosmic ray comes in, it's not really going to upset the veilance charge mechanism all that much compared to if there was a single like sort of point of failure, >> right? And so the there is the we have a >> thermal radiation benefit,

1:36:28>> a power efficiency benefit and a resilience benefit. And it's hard for me to imagine again there is the engineering problem but it's kind of like when Nvidia started with GPUs before you could really understand what the impact would be. We already can see now in the case of data centers in space why if you can figure out the industrial scaling of this >> it it it makes the execution of it in three the three arguably three of the most important categories yeah >> that you care about as an operator

What still needs to be solved

1:37:06>> significantly better. >> Yeah. in an order or orders of magnitude from what you would otherwise be doing with the current conventional options. >> Exactly. Yeah. I mean, I I think it's an incredibly exciting time for electronics at the cutting edge. Yeah. Right. Because all we're we're starting to see that things like GPUs and things that are capable of doing matrix multiplication and things like that are going to become increasingly important, perhaps even more important than traditional van architectures. Um, we've still got a long way to go. So, I do want to temper the brakes here. Um, and not dream too much because memory alone does not make a complete computer. You

1:37:47need high temperature logic circuits that also need to be developed and integrated alongside whatever chip that you're trying to put in. And current devices are built by hand submicron scale in the lab. The this is from their paper in the supplement. >> That length bar there says five micrometers. Okay. So that's that's like um a thousand times bigger than the chips that we have like the transistors that we have in our in our chips in my computer for example. So we got to make this thing smaller while still retaining the memorist characteristics the graphine that's going to be a problem right um scale up is going to take time

1:38:29on the manufacturing side. Two of the three devices, two of the three materials in this device, namely the tungsten and the halfneium oxide, those are already standard, right, in semiconductor manufacturing, right? Making the graphine smaller, that's going to be kind of tough, right? It's newer to the industry. But actually TSMC and Samsung both have that in their development road maps. If you go look, you know, they have like the shareholder like pitch decks for like we're still innovating guys, you know. Uh so they have graphine in their road map and it's already grown at wafer scale in research settings. We need to make it industrial if we want to

1:39:09have this thing grow. I mean the startup that um these USC scientists tetramm co-founded by Joshua Yang and Shia I mean it's already like doing things. So it we're we're we're heading for a new era in computation in a era where computation is becoming the bread and butter for our society more than ever more so than ever. I mean we used to think right ever since the advent of the '9s and the advent of the personal computer that computation would be big. Now it's like cloud computing and these large scale computing. >> Yes. >> Large scale computing is now taking precedence over personal small scale computing. Right. So, we're enter we're

1:39:50entering into a new era of computation. We've lived through one in the '90s and now this is a new one. >> As if uh as it relates to TSMC, it's interesting that both TSMC and Samsung are already working on you know, you know, integrating graphine into their industrial process. As if we needed another reason to for Taiwan to be a geopolitical, >> you know, uh matchbox. um particularly given the stakes of the perceived and real uh benefits of the race as it

Venus, geothermal, nuclear, and fusion

1:40:23relates to AI at large. >> Yeah. >> Um it this this is going to be and again you know we talk about breaking science research papers on this story. There is a delta between or a distance between it working in a lab and it being in your phone at home or in a data center in space. But it is a we're trying to continue to build this road map for people to understand where we're going as a society and what some of the best and brightest among us are trying to bring to us. And again we talked about this in the context of data centers in space but there are other things >> that you've mentioned you know that

1:41:05would be used this will be useful for in extreme computing planary exploration and other things. So >> space exploration I'd like to get better photos of Venus. >> Yeah. Yeah. >> That would be really nice. >> That'd be great. >> Um I'd like to have a rover on Venus that's like running around and stuff. That would be cool. um heavy energy like geothermal nuclear monitoring. I'd like to have safer nuclear >> y >> um power. >> If you're a green or clean energy person, I mean this can have real real impacts >> um in that space because it can help us better both understand and also execute in terms of production, maintenance, operational aspects.

1:41:46>> Yeah. Really anything that requires high temperature computing, right? I mean, with the advent of um fusion, right? If if we're trying to make fusion a thing, stuff is hot. >> Yep. >> And and so if we want um real time monitoring and decision-m using some AI algorithm about how to confine the plasma, I'm just thinking out loud, right? Like that's something >> that this would be used for. >> I don't need like to cool it down to room temp. I can just cool it down to 700°. >> Who needs a room temperature semiconductor? >> Yeah. >> Well, we still want that. >> We still want that. >> We still want that. Uh this was a really great electronics that can survive lava. Again, this was u a recent science paper

1:42:26March 2026, a collaboration led by researchers here in California, Forever Goodbye at USC, uh alongside the Air Force Research Lab in Dayton, Ohio, and the Kumamoto University in Japan with funding from the National Science Foundation, the Army Research Office, and the Air Force Research Lab.

Final thoughts

1:42:50uh can't wait to see uh uh high temperature memorers in a hypersonic vehicle near you. Um because certainly uh they're going to be using it for command and control systems integrated with, you know, next generation military stuff. So before they do that, we would like to have it be used for the good of mankind. Do you have any last notes before we wrap up today? No. Um, if you want to leave a comment, think of a really funny application for high temperature computing. Where would you want to put a chip? [laughter] >> Run wild. >> Keep it PG. Everybody, this is a familyfriendly show. Uh, as always, my

1:43:32name is Lester Nar joined by my co-host, our resident data centers in space are possible PhD Krishna Chowdery. We are approaching our oneyear anniversary. So, be sure to be tuned in for our anniversary episode. For our longtime listeners, we are extremely grateful. If you just happen to be listening to us for the first time, catalog is a bunch of episodes that are just like this. 50 of some of the best science entertainment on the planet. We will see you all next week.

1:44:12>> [music]

1:44:24[music]