Transcript
Stephen Cass: Hi there and welcome to Fixing the Future, an IEEE Spectrum podcast, the place we take a look at concrete options to some large issues. I’m your host, Stephen Cass, a senior editor at IEEE Spectrum. And earlier than we begin, I simply need to let you know you can get the newest protection from a few of Spectrum‘s most essential beats, together with AI, climate change, aynd robotics by signing up for certainly one of our free newsletters. Simply go to spectrum.ieee.org/newsletters to subscribe. In the present day, we’re going to be speaking about making tiny issues even tinier in order that we will cram ever extra computing energy onto silicon chips. And to try this, I’m speaking with one other Spectrumite, senior editor, Sam Moore, who covers a semiconductor beat for us like a field effect transistor masking a depletion layer. Sam, welcome to the present.
Samuel Ok. Moore: Thanks, Stephen. Nice to be right here.
Cass: So we’d typically discuss Moore’s law, no relation, on this present, and the present state of it and the way we at all times appear to be speaking about it’s the top of Moore’s regulation. And but, it retains going. So are you able to speak a bit about what the present state is and these new concepts for pushing that boat additional down the stream?
Moore: Positive factor. Sure, Moore’s regulation is slowing down. That’s a type of definitive truth. It’s getting more durable and dearer to make extra transistors on a given space of silicon. However it’s persevering with, and there’s numerous effort to make that occur. Proper now, we’re type of across the 200 million transistors per sq. millimeter vary, and going to proceed to maintain attempting to make that smaller. You hear quite a bit about type of five-nanometer node, three-nanometer node, and stuff like that. You need to needless to say these names really don’t have anything to do with the dimensions of stuff on issues. The five-nanometer node— sorry, the five-nanometer node chips usually have type of their smallest distance between the wires is 20 to 25 nanometers. So it’s all only a identify. However they’ll proceed to call new issues and make new processes and make issues even smaller.
Cass: So what do these names relate to in any respect then?
Moore: It’s historic. There was a time once they really meant one thing. That measurement, it’s referred to as the steel pitch. Mainly, the gap between two wires used to really be what they named issues after, excuse me. However that form of broke down within the late ‘90s or so. And so ever since then, it’s simply been type of a reputation.
Cass: And what’s it primarily based on?
Moore: Sorry. So what’s it primarily based on now?
Cass: Yeah.
Moore: Oh, properly, they stored slicing that distance in half and in half once more. After which they simply form of continued utilizing that type of division for the identify, despite the fact that it wasn’t really associated to the dimensions of issues on the transistor.
Cass: It’s identical to a common, “We’re getting smaller form of—”
Moore: Yeah, it’s simply form of advertising and marketing. And there’s so few firms who can really do these actually, actually cutting-edge chips, so they might name them something. And now they’re. I imply, TSMC now calls its course of N5. Intel goes to be— their subsequent era or the era after that will probably be like A20. I believe the A’s for Angstrom.
Cass: So for you, the actual measure, although, is that one you stated earlier, which is what number of million transistors you’re going to get per sq. centimeter sq..
Moore: Yeah. That’s actually what issues. It’s simply what number of you’ll be able to pack in.
Cass: So in the newest concern of Spectrum, we had a improbable characteristic that you just edited called the Moore’s Law Machine about a few of these efforts to type of preserve Moore’s regulation going with this fantastically elaborate gadget. And this was written by Jan van Schoot. So maybe you can take me by means of this expertise, which known as excessive ultraviolet, which appears like a soda I would purchase, however yeah.
Moore: Sorry. Properly, they might have referred to as it gentle X-rays. However excessive ultraviolet is unquestionably cooler. So excessive ultraviolet lithography is the way you make the— it’s the primary machine concerned within the newest two generations of cutting-edge semiconductor chips. So with out it, you wouldn’t have your iPhone 12 by means of 15, I imagine, the NVIDIA H100, that GPU that everyone’s attempting to get their palms on to do their AI. You wouldn’t have a minimum of one of many prime 10 supercomputers and possibly not one of the subsequent era of them. That is the crucial machine, and it’s made by one company in the Netherlands. And it’s fantastically sophisticated. Let me type of let you know first what it does, after which I’m simply going to provide you some bizarre superlatives about it. What lithography is, is type of, principally, you’ve obtained a sample that you just need to venture onto the chip that can ultimately make all of the circuits and transistors and issues like that. And with excessive ultraviolet lithography, you might be utilizing a wavelength of sunshine that’s solely 13 and a half nanometers lengthy. This can be a large soar from what was used within the earlier era, which was 193 nanometers, which was referred to as deep ultraviolet. It’s an unlimited soar. It took greater than 20 years of R&D to really get to a machine that works and that’s vaguely inexpensive. And once I say vaguely inexpensive, I actually imply vaguely inexpensive. Every machine is greater than $100 million. It’s obtained 100,000 or extra parts in it. It consumes a megawatt of electrical energy simply so it may ship a few hundred watts of this excessive ultraviolet gentle onto the silicon wafer.
The factor weighs like 180 tons. I imply, it’s huge. The present era is like trailer dimension. I noticed certainly one of them being put collectively at a fab in upstate New York about 5 years in the past. And it’s so large everyone appears like an Oompa Loompa. I imply, very. If the Oompa Loompa’s had been the most effective on the planet at what they did and the chocolate manufacturing unit price a billion {dollars}. It’s simply a tremendous machine. And the following model of this machine, which is type of what we’re going to speak about at the moment, is greater than a 3rd bigger than at the moment’s. So it’s simply huge, sophisticated, tremendous costly, exhausting to get your hand on. And I need to let you know how they really made it higher.
Cass: So inside this large trailer machine, there’s some actually loopy parts, together with how they really make these gentle X-ray/excessive ultraviolet blink beams. And it entails molten steel and a carbon dioxide laser.
Moore: Yeah, it’s the most bananas course of you’ll be able to form of consider. So that you’ve obtained a vacuum chamber and a bit— I don’t know what to name it, however it’s spitting tiny molten drops of tin. And so they shoot throughout the vacuum chamber, and then you definately hit that tin along with your 40 kilowatt laser. You blast it right into a plasma. After which this plasma glows in all types of improbable colours. However the optics accumulate the 13 and a half nanometers that you just really need to use and venture it into the heart of the machine itself.
Cass: This actually does sound very Wonka-ish.
Moore: Yeah, it looks like there must be a neater manner to do that, however apparently, there isn’t.
Cass: So these items should be vacuum-sealed as a result of this gentle will get absorbed by the air. And what are a few of the different— why is that this machine so large? As a result of it looks as if you’ve obtained like a bit tiny sample. You’ve obtained little tiny chips, okay? These little tiny droplets. Why is it so large?
Moore: Properly, numerous it’s really the chamber containing the optics, that are simply insanely exact. The mirrors are fantastically costly. These aren’t simply bizarre mirrors. They’re a number of layers of alternating unique stuff to be able to get this type of gentle to mirror in the suitable route with any effectivity as a result of UV is absorbed by tons and tons of stuff, together with air. And so numerous it’s simply getting the sunshine the place it must go effectively with out disturbing any of the patterns that you just really need to venture. After which there’s numerous it that’s additionally simply dealing with the wafers and dealing with the masks, which once they’re really type of in place, they obtained to be dealt with at nanometer precision. So these are extremely fantastic shifting machines.
Cass: So one of many large challenges I believe that Schoot talked about within the article was, “Sure, you’ll be able to set this up, and you will get the lasers going, the machine’s going, however your throughput goes to be very, very, very low, uneconomical for the dimensions of the machine except you attempt a few different tips on prime of the, once more, molten droplets being blasted by a really highly effective carbon dioxide laser,” which I do know I’m hung up on, however inform me a bit bit extra about these different tips.
Moore: Positive. Okay. In order you’re a little hung up on— one of many largest issues they needed to resolve simply to get to the primary era was to make the tin explode in additional brilliance in order that you can get simply that couple hundred watts of sunshine as a result of, the dimmer it’s, the longer you must expose the wafer. And so it’s all about throughput. That downside is principally solved. However to be able to proceed Moore’s regulation, you don’t simply need this. You’ve obtained this gentle, however you need to really preserve making smaller and smaller options with this gentle. To do this, there’s three knobs you can flip. One was the large knob of adjusting the wavelength, which is sensible. Smaller gentle, higher decision, completely simple. Two different knobs. One is form of tough to clarify. It’s a bunch of optical tips that you are able to do, which could embrace as a lot as projecting two patterns serially to get one sample on the finish or simply making issues look bizarre in order that they give the impression of being much less bizarre once they get to the silicon wafer. And I can speak extra about that later.
However the knob that they’re turning with this latest machine that’s at present being inbuilt Belgium proper now could be to extend the numerical aperture. That’s type of the angle of the sunshine you can function inside the optical system. Traditionally, they’ve turned all three knobs. Numerical aperture is one that really offers you a extremely good return, traditionally. And they also actually wished to do that. They’re proper round 13.5 nanometer decision now, but when they need to dip down under, they’re going to should do a excessive numerical aperture, excessive ultraviolet lithography. This causes a cascade of issues once you’re designing the system. This machine is already fantastically sophisticated. However as with all optics, you tweak one thing right here, it’s going to have another impact afterward. So let me undergo the cascade of the issues that they needed to resolve to be able to make excessive NA excessive ultraviolet lithography.
Okay. So first, you need the numerical aperture improve on the wafer itself. That’s the place you’ll get the decision. However which means you even have to extend it on the masks. Now, the masks is the place you retailer the sample. So that you bounce the sunshine off the masks and goes by means of the optics, after which it lands on the silicon wafer, and that’s your level. Right here’s the factor. So you bought to bounce the sunshine onto the wafer— sorry, onto the masks after which off of the masks. And right here’s the place you’ve got type of a Ghostbusters second. These two streams can’t cross. You suppose unhealthy issues will occur. It would be bad, I believe, is the road, proper? So you’ll be able to’t cross the streams. So which means you must angle them away from one another, okay? However you’ll be able to’t angle them away from one another an excessive amount of as a result of these actually specialised mirrors solely work as much as about 11 levels. And to be able to simply— for those who simply wished to do that with none changes, you’d want 18 levels. So then they’re like, “Okay, properly, now we’ve obtained to unravel. We’re going to wish this angle. There’s no getting round it.” So the way in which they solved it was by growing the demagnification, which I do know sounds form of growing a D, however principally, it’s shrinking stuff down quite a bit. They elevated it by like eightfold or one thing like that.
So like, “Okay, hey, downside solved.” However probably not as a result of now your sample on the wafer is admittedly small. It’s like a postage stamp as a substitute of— not your sample, sorry. The quantity of wafer you can venture onto all of sudden, tremendous small. It’s like a postage stamp, and that implies that you must do extra postage stamps, which implies that you’re—
Cass: As a result of these wafers are giant.
Moore: Yeah, they’re 300 millimeters throughout and so dinner plate dimension or so. And so for those who can solely perform a little bit at a time, it’s going to take you longer to do a wafer. After which it principally turns into so costly, it’s not even price it. So that they needed to resolve that downside by doing one thing form of bizarre. It was type of like form of funhouse mirror results. Mainly, they elevated the demagnification in just one route. So that they got here up with these specialised mirrors that form of would stretch issues out [laughter] and shrink them. And it had bizarre results. I imply, you really should make the masks stretched out. So the sample that’s on the masks is form of this funhouse mirror model of what you need on the wafer. However amazingly, that really does it. You continue to wind up with a bit bit smaller than you’d like of a projection onto the wafer, however it’s acceptable so long as you improve the acceleration of how briskly issues are shifting by means of the machine. So what? 5 issues to unravel?
Cass: Yeah, I believe so, yeah. Yeah. However this impact, it sounds a bit bit like within the previous days earlier than all of us had widescreen, flat-screen TVs, generally once you had been attempting to be exhibiting like a cinema film on a TV and all of the sudden, the facet ratio would get actually bizarre and distorted as a result of they had been having to squeeze in on one axis to make all of it match, particularly when the credit would roll and would get all distorted in a single route. And that form of jogs my memory of that. However these are large machines. And also you’ve achieved another reporting, although, on a few of the type of the unwanted side effects of dealing with these large machines, which is function them in a extra type of environmentally pleasant manner. And that was the work of this firm referred to as, I believe, Edwards in England. So are you able to inform me a bit bit about that?
Moore: Yeah. So that you bear in mind how I stated the whole lot has to occur in a vacuum on the within of this? Form of type of not. There’s a very small type of circulate of— I imply, sorry. It looks as if a small circulate of hydrogen, however it’s a extremely large machine, so it’s really 600 liters per minute. [laughter] However this hydrogen is there for a few causes. Every little thing in there’s tremendous delicate. You don’t need something to get on the mirrors or on the masks or something like that. However you might be blasting molten tin in a chamber, and you’ve got different chemical substances which are concerned in chip making and stuff like that, and you might want to form of sweep them away. And in order that’s what this hydrogen is for. And also you suppose, “Oh, hey, hydrogen. That’s inexperienced.” Not but, really. Most hydrogen is definitely not made in any inexperienced course of. It’s really made by a chemical response between water and methane, so not nice. And 600 commonplace liters per minute is definitely form of quite a bit. What they’re doing with it at present is they simply burn it since you simply get water, and the entire nasties that it’s picked up simply form of falls out. And that’s—
Cass: Form of a smoke stack on the facet with a flame on prime, [laughter] and it’s much more Willy Wonka. However yeah, okay. So that they’re simply burning off the hydrogen.
Moore: Proper. However that’s tremendous wasteful. So what Edwards labored on was a system that may recycle the hydrogen. It’s really fairly cool. It’s like a reverse gasoline cell form of. The used hydrogen and the icky parts that it’s picked up alongside the way in which. Mainly, undergo one facet, get ionized. Then an electrical subject type of forces these protons by means of a proton change membrane. They arrive out of a membrane, excuse me. They arrive out the opposite facet, recombine with electrons. You get pure hydrogen to return into your course of. And all of the terrible stuff stays on that to the opposite facet. Yeah. So it really works fairly good. They set one up at a nanoscience analysis group referred to as IMEC, which is form of a key European analysis home. And it managed to recycle 70 to 80 % of the hydrogen of their EUV machine. So now they simply should persuade the large chip makers to undertake it as properly.
Cass: So with all of those applied sciences— and I need to flip to type of a competing expertise in a second, however for these applied sciences, how lengthy do you suppose it’ll be earlier than we see chips in our smartphones and our computer systems made with this new expertise?
Moore: Proper. With a excessive NA EUV, issues which are made in 2025 will most likely begin to— a minimum of the chips themselves will probably be made in 2025. It takes months after that for them to be in programs, however that’s most likely in time for yet one more NVIDIA GPU. So AI is driving numerous the demand for significantly this most innovative. And so I’d anticipate to see it in type of the era of AI chips which are type of made in 2025, ‘26. Additionally, Apple is at all times on the innovative. They at all times need the latest chip manufacturing methods. So no matter iPhone [laughter] comes out within the latter half of the century will virtually definitely contain this.
Cass: And that’s really an ideal segue since you talked about NVIDIA there. And NVIDIA is taking a look at enabling one other method to squeezing issues down and maintaining Moore’s regulation shifting alongside, which is inverse lithography. So are you able to inform me a bit bit about that? And why the truth that it’s an AI firm works out for properly for them? As a result of they’re a chip maker who occurs to make just like the AI chips.
Moore: Proper. So let me type of give us a bit extra context because you began. So NVIDIA, really, they design essentially the most in-demand AI chip on the planet. All people desires their palms on an H100, which is simply the present era. The producer of that chip, although, NVIDIA designs it. It’s manufactured by TSMC, which frankly, form of dominates the world of essentially the most cutting-edge chips proper now. So that they work intently collectively now as a result of NVIDIA might be certainly one of their largest prospects. So for those who can form of return for a second, bear in mind I instructed you about these three knobs you’ll be able to flip to make lithography higher, to make your precision and your decision higher. A kind of knobs was this bizarre one referred to as K1. It was type of the method stuff that you are able to do. So what NVIDIA has achieved is it’s made a kind of course of stuff that you are able to do a lot simpler to compute. It’s a method referred to as inverse lithography. And right here’s the factor. You would possibly suppose that for those who wished to venture, say, like a plus signal, one thing that was formed like a plus signal onto a silicon wafer, in your masks, you’d put a plus, and then you definately’d get a smaller plus when it obtained to the silicon. And never the case. There’s sufficient distortions and different simply stuff that you must fear about optically once you’re coping with this type of operation under the wavelength of sunshine that you just’re utilizing, that you must do issues like add little type of canine ears on the finish of the plus signal to make it appear like a plus when it will get there. These issues have needed to be progressively extra sophisticated as we’ve form of pushed Moore’s regulation to its limits.
So now that plus signal would really type of appear like— for those who put it in a kaleidoscope and form of turned it, it’s simply this huge bizarre stuff that you must placed on the masks to be able to get your easy plus signal on the wafer. Now, these tips are literally form of actually exhausting to do computationally. So it’s the concept that like, “Okay, if I need this plus on the wafer, what do I’ve to have on the masks?” And it’s so computationally tough that we’re speaking like weeks of simply— we’ve obtained a large laptop, and it’s going to only sit there for a few weeks and check out to determine what that form must be. What NVIDIA has achieved is it’s give you a system that turned that two weeks into an in a single day job. And the factor that was that it was— it was a job for CPUs. My guess is it was instructionally sophisticated sufficient that it was not type of inherently of the parallel nature that GPUs had been on. So NVIDIA did numerous work and got here up with algorithms which are simply completely fitted to a GPU. And so, principally, they did within the work— sorry, what would have taken 40,000 CPUs they did with 500 GPUs and two weeks versus in a single day, which is definitely— that’s large. That eliminates a giant bottleneck in getting your chip to market, for one factor. It permits you to use this actually computationally costly method in additional locations somewhat than reserving it for the spots of the chip that had been simply actually tough.
And from the angle of an environmental profit, it’s 5 megawatts of energy within the computing system versus 35 megawatts, which isn’t insubstantial. So yeah, it is a factor that they— this computational lithography system, it’s referred to as cuLITHO. They launched it, I believe, within the early summer time or late spring, and so they’ve obtained Synopsis, one of many digital design automation firms purchased in. TSMC has been engaged on it with them. And naturally, ASML, which makes the EUV machine in query. And all the opposite lithography machine gear they’re signed on as properly. So it ought to actually be making a distinction each environmentally and by way of getting chips achieved quicker.
Cass: So simply to wrap up, we’ve been speaking about numerous applied sciences which are really very near being deployed. Is there something you’re seeing within the lab that’s additional out which may assist us like within the 2030s, principally?
Moore: There’s no clear reply as to whether there’ll be type of one other wavelength of sunshine that we use, and it appears form of unlikely, really. Even 13.5 nanometers will not be that many atoms of fabric once you get right down to it. So our means to type of shrink issues down in two dimensions, it truly is getting in the direction of the top. And so transistor structure is beginning to go 3D. Or somewhat, within the lab, it’s beginning to go 3D. However this looks as if the trail that everybody has chosen. So now there’s a brand new form of transistor. I imagine Samsung began utilizing it in manufacturing final 12 months, perhaps TSMC this 12 months. I might need these improper, however they’re each properly into this new construction. It’s called a nano sheet. And Intel is shifting to it the top of 2024. And the factor concerning the nano sheet is that it’s type of conducive to creating a second transistor proper on prime of it. So as a substitute of attempting to squeeze issues collectively in two dimensions, we’re going to start out including layers. Along with simply on the transistor stage making it 3D, we’ve already obtained numerous work occurring proper now and numerous manufacturing chips that contain 3D packaging, which is simply taking one chip and stacking it on prime of one other to be able to form of make a superchip. And that’s taking place now in manufacturing chips. So yeah, the longer term is three-dimensional.
Cass: Properly, that was improbable, Sam. Thanks a lot for speaking with us at the moment.
Moore: It was a pleasure, Stephen, as at all times.
Cass: So at the moment we had been speaking with Sam Moore, senior editor at IEEE Spectrum about excessive ultraviolet and different applied sciences to maintain transistors getting ever smaller on laptop chips. For IEEE Spectrum, I’m Stephen Cass, and I hope you see us subsequent time on Fixing the Future.