Semiconductor companies are pouring large sums into ASML, and ASML is pouring large sums into its suppliers. CTO Martin van den Brink on how EUV lithography has changed the rules of the game, and on the next generation of EUV scanners.
Martin van den Brink is paging through the previous Bits&Chips ASML special as his visitor is ushered in. He leafs to the interview he gave for that issue. ‘Let’s see what I said last time,’ he grins once the handshakes are over, and he reads off a few arresting pull quotes: ‘‘The source is not a pretty story and it won’t become one anytime soon’, ‘Cymer has great ideas about LPP, but the execution is very disappointing’, ‘EUV will remain a challenge for the next decade’. And not a word of it wrong,’ he says triumphantly.
We encounter the company’s president and CTO – technology, strategy, marketing: what doesn’t he do? – in excellent spirits. But not for the reason we’d expected, our conversation soon makes evident. Van den Brink doesn’t want to crow too loudly about the fact that his customers are now openly announcing plans to incorporate ASML’s EUV scanners into their production processes – a painful ten years later than the company had intended, but nonetheless a giant leap compared to the situation two or three years ago.
‘Our customers were starting to despair,’ he says matter-of-factly. ‘Multipatterning is really going to start claiming a significant chunk of their budget now. When double patterning first became necessary, we were talking about two or three layers. Now we’re dealing with dozens. They’re getting desperate; it’s as simple as that. Of course our machines’ performance is also improving, but that doesn’t mean they’re perfect.’
The evidence for that is lying right on his desk: a small plastic tube filled with silver-grey blocks and two small machinist’s squares lined with a solder-like substance. ‘This came out of the source,’ he explains. ‘We’ve gotten to the point that we can keep the collector pretty clean, but the tin still ends up in places where it shouldn’t.’ There’s still a great deal to be done, in other words, not least because EUV sources must eventually match the performance of 193-nm lasers. ‘We’ll been working on that for some time to come,’ Van den Brink declares.
Van den Brink’s good mood does turn out to be related to EUV after all – or, rather, to its future. He feels that a corner has been turned. Not only at his company, but also in ‘public’ opinion. He himself has never doubted that the EUV project would turn out well, but he’s shared lunch with a large number of people who have. ‘These last few years, I’ve spent a lot of time shoring up people’s self-confidence, convincing them they really can do it.’
The pep talks have had the desired effect, as evidenced by the deal announced two weeks ago with the semiconductor division of Carl Zeiss. ASML is putting up one billion euros for just under a quarter share in its primary optics supplier, plus 760 million euros paid over six years to intensify R&D efforts and for other necessary investments related to developing the next generation of EUV machines, what’s known as high-NA scanners. ‘The fact the investment went through – here and at Zeiss – is the greatest vote of confidence you could have,’ Van den Brink says.
‘If I ask Peter [Wennink, ASML’s CEO] for two billion, when not a single engineer here believes in high NA, well, I can just forget it. When I first brought it up three or four years ago, everyone said: “Come on, Martin; we’re already so tired.” Now they see it differently; no one’s criticizing it now. Not in the outside world, either. Last week I spoke at our analysts’ meeting in New York and there, too, our plans were received with enthusiasm.’
With high NA on the roadmap, he says, the company knows what it will be doing for the next ten years. ‘We have a clear course and strategy, one that young people can run with. That gives me an incredibly good feeling, and I hope it gives our employees one, too.’
Why is this kind of arrangement with Zeiss necessary?
Van den Brink: ‘I’ve been saying for thirty years that the way we’ve partnered with Zeiss isn’t sufficient to reach the finish line. Most lithography companies already understood lenses; the technology was comparable for a long time. But that’s absolutely no longer the case today. Zeiss knows that everything it develops for EUV comes exclusively to us. What’s more, in thirty years’ time the cost of a light source has gone from a few thousand euros to an eight-figure number, a substantial share of the cost of a machine. That makes both of us vulnerable, and that’s why we have to take joint responsibility for the necessary investments. We have to redefine the customer-supplier relationship.’
Was there resistance at Zeiss?
‘Zeiss is a different kind of company than we are. It isn’t publicly traded, likes to make its own way and prefers not to be dependent on others. It takes time for management to develop trust in a new type of partnership. And of course the deal involves the question of what a lens like that should cost. There has to be something in return for our investment.’
Did the acquisition of Cymer in 2012 have a similar motivation?
Van den Brink nods. ‘The mercury bulb we used to have in our steppers wasn’t fundamentally different from the bulb you’d find in a street light. But Cymer’s lasers were only going to a very select group of clients. So you’re working in nearly the same one-on-one relationship as with Zeiss. Meanwhile, the light source has grown to encompass a large part of the system and heavy investments are required for the future, which Cymer itself couldn’t make.’
Yet the impression arose that Cymer couldn’t do the job, and that’s why you acquired it. In fact, you said so yourself in the previous interview.
‘Look, today 60 per cent of the work on the source is being done in Veldhoven. We scaled that up. So the whole thing was underestimated. Most of that scale-up is in engineering. If you look at the fundamental research, a large part of the technological progress is being done in San Diego. Turning that progress into a machine that has to perform day in, day out at the customer: that’s where we’ve made real strides.’
So did you wait too long to acquire Cymer?
‘Yes. About fifteen years ago, I was already talking with Willem Maris [ASML’s former CEO] about the EUV source. I told him, ‘We’re going to do EUV, but we’ll never get the source right if we don’t take on the responsibility for it ourselves. That thing is way too complicated.’ I suggested to him that we set up a joint venture.’
But that was a different era. ‘Back then, the thinking was: if you don’t have to, you don’t annex suppliers’ expertise. You look around and choose a partner who does the best work. That’s still true, except when you’re talking about unique expertise that has no other application. A few years ago that wasn’t all that clear, but the rise of EUV has changed that. And then it took some time before the minds in San Diego were ready for it. We can’t afford to take an aggressive tone; every decision requires thorough discussion, and that takes time. And Cymer was a publicly traded company; that makes it more complicated.’
In 2012 Intel, Samsung and TSMC took a stake in ASML and started helping to pay for R&D. Are you developing the same kind of relationship with these semiconductor manufacturers as with Zeiss and Cymer?
‘The dependence on each other is the same. You can’t avoid that. The technological challenges require that we work closely together and reach high-level agreement on things, but in a way that each party is comfortable with. Just this week we sat down with Samsung’s management and had a very amiable discussion. The CEO was there.’
Can we expect more of these kinds of deals in the future?
‘Not at this scale, if they happen at all.’
By jumping in wavelength from 193 to 13.5 nanometres, EUV scanners can image much smaller details than their excimer-based predecessors. But not proportionally smaller ones. Because in addition to wavelength, the numeric aperture (NA), or lens opening, has just as much inverse influence on maximum resolution, and EUV can’t achieve the same NA that laser light can. ASML’s newest 193-nm Twinscans have an NA of 1.35; the current generation of EUV scanners, 0.33. There’s also a third factor, the k1 coefficient, but Van den Brink says they’ve squeezed all they can out of that.
The gain provided by the jump in wavelength is nonetheless substantial: in a single exposure, an immersion scanner can print 40-nm structures, and no smaller. An EUV scanner can go down to 13 nm. This resolution is enough, Van den Brink estimates, to meet the requirements of the 5-nm node. The International Technology Roadmap for Semiconductors positions this generation of chips between 2020 and 2021. At current speeds, Intel – whose definition is closest to that of the ITRS – will probably take longer to get there, while the foundries – who mix more marketing into their definitions – will already be making 3-nm chips by that time.
‘After 5 nanometres we’ll have to use double patterning again, but this time with EUV,’ Van den Brink says. ‘If we can raise the NA from 0.33 to 0.5 or maybe even a little higher, we’ll gain nearly a factor of 2 in resolution and we’ll push double patterning back two generations.’
The step to an NA of 0.5 will be a tough one, however. Not least because ASML doesn’t want to saddle the EUV ecosystem with new changes to the infrastructure. The mask, the resist and the pellicle – each of them elements of EUV lithography that still call up concerns – must remain compatible with current machines. What’s more, the scanners must not become slower.
Van den Brink describes the challenges. ‘If the aperture increases, the angle at which the light hits the mask and is reflected increases. Because mask patterns are three-dimensional, shadows start to cause problems. What’s more, the angle affects the performance of the resist. That’s made of several layers of coatings with different refractive indices, so that the spacing of the layers resonates constructively or destructively with the wavelength. That trick only works at one specific angle, and our angle will change.’
‘To make a long story short: we have to limit our angle. We can do that by choosing a larger magnification, but that has the consequence that the dimensions of the exposed field become smaller. Now the range of angles is only critical in one direction, because the incoming and outgoing bundle are only kept apart in one direction. If we increase the magnification in one direction from 4x to 8x and keep the other direction at 4x, we can use the same size mask to make chips that are only twice as small as the current size.’
What ASML does by using different magnifications in different directions is comparable to a technique from the film industry: the anamorphic format. Using an anamorphic lens, a widescreen picture can be shot on standard film, and anamorphic optics enable ASML to make its exposure field twice as small in the high-NA machine, instead of the 4x it would get if the magnification were doubled in both directions.
That’s important, because if the field is four times smaller, the machine has to do four times as many exposures to print a full wafer. Thanks to anamorphic optics, the damage is limited to a factor of 2, but Van den Brink’s throughput still takes a hit. And the angle problem still hasn’t been solved, either.
It’s too complex to explain in detail, but that can also be fixed, by using catadioptric optics. In photography, catadioptric lenses are a combination of regular lenses and curved mirrors, characterized by a central obstruction. They used to be sold as cheaper and lighter-weight telephoto lenses, but their popularity suffered from an unavoidable out-of-focus torus in the resulting image, among other things. No problem, says Van den Brink. ‘Our software will calculate that out.’
For ASML, catadioptric optics boil down to mirrors with a hole in the middle. That results in the loss of some light, but the range of angles is decreased. What’s more, because more ‘useful’ light reaches the wafer at these angles, an exposure can be carried out more quickly and thus the machine can process more wafers per hour. In New York, Van den Brink presented the attending analysts with an improvement in throughput of roughly 25 per cent compared to 0.33-NA EUV scanners.
‘Of course, we have to scan twice as fast, and that’s an enormous challenge for our staging team. But that’s absolutely our core competency. And we’ve got the best people in the world,’ the CTO says as he puts down the pen with which he’s sketched diagrams for his visitor. ‘My people are very enthusiastic about this. It energizes them. There are very few companies in the area who can tell you what they’ll be doing for the next ten years.’
The obvious next question is: how small can you go?
‘Lithographically speaking, we don’t see a theoretical limit. If a customer sees an opportunity, we’ll help them create it. It does get increasingly complex, but technologically, the end won’t be determined by us. Going to an even higher NA is pointless, but we can go down another factor of 2 in wavelength. Then the infrastructure will have to be overhauled.’
Yet the papers are full of reports that Moore’s Law is dead. Intel has slowed down.
‘There’s absolutely no reason to assume that Moore’s Law will stop. Other than that several people think it’s getting very, very complex. But no one has slowed down substantially. So saying that it’ll all be over tomorrow, that’s completely unfounded. Fundamentally, nothing has changed.’
Is it conceivable that geometric scaling will decline in importance, in favour of other methods to improve chips’ performance, such as 3D integration?
‘Geometric scaling is indeed one of the engines that drives Moore’s Law. There are also integration, device scaling and architecture. I don’t see them as opposing forces. All four will keep happening. As Mike Mayberry at Intel taught me: as long as there are new ideas, the roadmap will not stop.’