Kind: captions
Language: en
This is a microchip. When you zoom in,
you find a nanoscopic computing city.
Skyscrapers hundreds of layers tall with
hundreds of kilometers of wires
connecting everything. And at the very
bottom is this transistors,
billions of them. They are the ones and
zeros of our computer. The chip works by
whizzing electrons from transistor to
transistor. And the smaller you can make
those transistors, the less the signals
have to travel. So the faster they can
compute. Plus, you can fit more [music]
transistors into the same area,
resulting in a much more powerful chip.
So, for over 50 years, transistors got
smaller and smaller, and the number you
could fit on a chip doubled every 2
years. This became known as Moore's law,
named for Intel's co-founder Gordon
Moore after he noticed the pattern back
in 1965, and it's been one of the main
drivers of the tech industry. But around
2015, progress [music] came to a
screeching halt. And we might have never
gotten past it if it wasn't for a single
company that makes these machines. The
machines that saved Moors law. Holy.
This is a video about the most
complicated commercial product
humanity's ever built. That's insane. It
costs a whopping $400 million. And it is
so bizarre that I want to introduce it
to you with a thought experiment.
Imagine you are shrunk down to the size
of an end and you're given a laser
that's strong enough to melt through
metal like butter. Next, a tiny droplet
of molten tin, roughly the size of a
white blood cell, is shot out in front
of you around 250 km hour. And your task
is to hit this not once, not twice, but
three times in a row in 20 micro seconds
with your little laser. Well, that is
exactly what this machine [music] does.
It hits one tiny tin droplet three times
in a row, heating each one up to over
220,000 Kelvin. That's roughly 40 times
hotter than the surface of the sun. And
it doesn't just hit one droplet, it hits
50,000 droplets every single second.
>> How often do you miss a laser shot?
>> We don't miss them.
>> What? You do 150,000 laser shots a
second and you don't miss one.
>> Exactly. The same machine also contains
mirrors that might just be the smoothest
objects in the universe. If you scale
one up to the size of the Earth, then
the largest bump would be no thicker
than a playing card. On top of that, it
is able to overlay one layer of a chip
perfectly on top of another and never be
off by more than five atoms. And this is
all happening while parts of the machine
whip around at accelerations of over 20
GS. For 30 years, almost everyone
thought that actually building this
machine was impossible. And yet, it
exists. There is only one company in the
world that can make it. So, what is this
company? And what is this impossible
machine they've built? This video is
sponsored by Brilliant. More about them
at the end of the show. Now, just as a
quick aside, the makers of this machine
didn't actually sponsor this video. We
just thought that the science and
engineering here were so cool that we
had to make a video about it. So, let's
jump straight in. To make a microchip,
you start by taking silicon dioxide,
usually from sand, and purifying it into
ultra pure, nearly 100% silicon chunks,
which is then melted down in a special
furnace. Next, you lower a small seed
crystal into the vat. Silicon atoms
attach to the crystal, extending its
structure. Then you slowly raise the
seed crystal while rotating it. And this
results in a large single crystal
silicon ingot.
>> This is where the seed crystal would be
>> and then you pull it out.
>> Can I touch it?
>> Yeah, you can.
>> It seems like you would not be able to
hold this from here.
>> Yes.
>> It even feels fragile like if you kind
of
>> Yeah. I'm scared to break it.
>> Yes.
>> He's using more force.
>> The ingot is then cut into wafers with
diamond wire saws. Up to 5,000 of them.
after which each wafer is carefully
polished. Next, it's coated with a light
sensitive material called photoresist.
[music]
There are different kinds, but in a
positive photoresist, the areas exposed
to light become weaker and more soluble.
So, if you shine light through a
patterned mask, you can selectively
weaken parts of that coating. [music]
Then, you rinse the wafer with a basic
solution to wash away the exposed
photoresist, leaving the design
imprinted. [music]
So, now you can actually turn this
pattern into physical structures. This
is often done by etching into the
uncovered silicon by using either
chemicals or plasma. And then you
deposit a metal like copper to fill in
those etched lines. As a last step, you
wash away the remaining photoresist. And
now you've made a single layer of the
chip. We've simplified this cycle down
to the main steps. Coat, [music] expose,
etch, and deposit. It repeats for every
single chip layer. And depending on the
chip, there could be anywhere from 10 to
100 layers. The bottom layer is the
[music] transistors. This is the most
complicated layer, requiring hundreds of
steps that all need to be perfect. The
higher layers are a little easier. These
are the metal wires that carry signals
and power. By the end, the completed
wafer can have hundreds of chips,
[music] which are then cut into separate
pieces, packaged, and put into products.
But by far the hardest and most crucial
step in the process is where you shine
light through the mask and onto the
wafer. This is photoiththography. And
that's because this step determines how
small you [music] can make the features.
At first, it seems simple. Light passes
through the openings and it gets blocked
by all the rest. But as you try to print
smaller and smaller features, the gaps
in the mask start to approach the
wavelength of the light. [music] And
that causes problems. And we can
actually show it because I happen to
have a uh this is a mask. This is a
reticle. A reticle or a mask carries the
design of one chip layer. This reticle
is filled with microscopic lines and
gaps around 670 nanometers across.
>> And if I take like a laser pointer, so
this is a red laser.
>> Yeah.
>> If I shine it through it, then you see
this here.
>> The laser has a wavelength of around 650
nm. When light hits the reticle, its
wave fronts bend as they pass through
each gap. So each gap sends out waves
that spread out and overlap. Now let's
just look at the light from these two
gaps. When the peaks of one wave line up
with the troughs of the other, we say
that the two waves are out of phase and
they cancel each other out. So you get
dark spots. And when the peaks line up
with the peaks, the two waves are in
phase. They add up and you get bright
spots. [music] You get interference,
right? And you get a defraction pattern.
>> Now defraction is inevitable. So instead
of fighting it, designers actually use
it to get the patterns they want. They
kind of work backwards from the eventual
pattern they want on the wafer and they
design the slits so that defraction
[music]
will occur in such a way that it creates
the pattern that they want. You see
three dots. Uh the middle dot that's the
original one. That's the zero [music]
order. And then on the left and the
right you can see the first and the
minus first. Now in order for [music] us
to have this image resolved on the wafer
you need to capture the zero and the
first and the minus first [music] order.
The smaller you make the features, the
larger this angle alpha between the zero
and first orders becomes. So the larger
your lens needs [music] to be to capture
the light. The size of the lens is
described by the numerical aperture or
NA for short, which is just the sign of
this angle. So the larger that is, the
smaller the features you can print. But
there is a hard limit to how large your
lens system can be. When this angle is
90° and your numerical aperture is one,
well, your lens would have to [music] be
infinite. Fortunately, there is one
other thing we can change.
>> This is a red laser.
>> Yeah.
>> And a red laser has a wavelength of 650
nanomeish, I would say.
>> And if I take a green laser, and this
one has a wavelength of uh 532, then you
can see that the green dots
>> are closer spaced [music]
>> than the red dots.
That's because the light from the two
different gaps doesn't have to travel as
far to match up in phase again. So the
orders end up closer together. So with a
smaller wavelength, you can print
smaller patterns using the [music] same
lens. All of this is captured by the
equation which determines the smallest
feature size or critical dimension.
But since there's a limit to how much
you can increase the numerical aperture
I mean two one over time the only way to
keep making smaller and smaller features
is by using shorter and shorter
wavelengths. [music] So this is exactly
what happened up until the late 1990s
when the industry settled on 193
nanometer deep UV light. This was the
light that was used to make all of the
most advanced chips right until around
2015. But by that point, scientists had
reached the limit to how small they
could make the features. And Moore's law
was about to run into a brick wall. So a
radical change was needed. A change that
had been brewing for around 30 years.
All the way back in the 1980s, Japanese
scientist Hirokino came up with a crazy
idea. Why not use much shorter
wavelengths like X-rays of around 10
nanometers? In theory, that should allow
you to print much smaller features. But
you quickly run into a problem. X-rays
at these wavelengths have enough energy
to eject electrons from their atoms. So
most materials absorb them. But unlike
medical X-rays, which have wavelengths
shorter than 1 nanometer, these are
still long enough to interact with air.
So air absorbs them, too. That meant
that Kinoshida's setup had to be in a
vacuum. But even worse, he couldn't use
lenses to focus the light because the
lenses would absorb it too.
So it seemed like this idea would never
work.
But around 1983, Kinesita stumbled on a
paper by Jim Underwood and Troy Barbie.
Their work focused on special mirrors
that could reflect X-rays with a
wavelength of 4.48 nm. So Kinesisha was
intrigued. Curved mirrors can focus
light just like lenses do. If he [music]
could figure out how to make these
special mirrors for the wavelength he
was using, then this could be another
way to do photoiththography. The [music]
mirrors work something like this.
When light crosses from one medium to
another, say from air to glass, it bends
or refracts. Some of it goes through and
part reflects back. How much gets
reflected depends on things like the
angle, the light's polarization, and
most importantly for us, the difference
between the refractive indices [music]
of the two media. The larger that
difference, the more light is reflected.
And Underwood and Barbie used that
[music] principle. They made a super
thin layer of tungsten less than 1
nanometer thick. Thin enough that X-rays
could pass through without immediately
being absorbed. When X-rays hit the
layer at a specific angle, the tungsten
reflected less than 1%. Then they
carefully tuned the layer thickness so
the path length of the transmitted
X-rays was only one quarter of its
wavelength. Then they added another
layer, this time out of carbon. It has a
higher refractive index [music] than
tungsten for wavelengths of 4.48 nmters.
The X-rays hit the boundary and a little
bit more reflects. But this time, the
phase is inverted or it's changed by
half a wavelength. This happens when any
light moves from a lower refractive
index to a higher one. Now, by the time
this new reflected wave reaches the
tungsten boundary, it has traveled
another quarter of its wavelength for a
half wavelength in total. So the two
phases line up and the waves interfere
constructively. Underwood and Barbie
kept doing this trick for a total of 76
alternating layers so that in total they
could reflect back much more of the
X-rays.
Now they only managed to reflect around
6% of the light but it was a proof of
principle that you could reflect X-rays.
So Kinoshida saw the possibilities. He
got to work and after around two years
his team designed and built three
tungsten carbon curved multi-layer
mirrors to reflect 11 nanometer light
and with it he managed to print lines
four microns or 4,000 nm thick proving
that at least in theory X-ray litography
was possible.
A year later in 1986 he went to present
his findings to the Japanese Society of
Applied Physics. Proud and excited he
explained his setup and showed his
image. But to his horror, the audience
refused to believe it.
>> Unfortunately, the audience [laughter]
was highly skeptical of my talk.
>> Kinoshida was devastated. He later said,
"People seemed unwilling [music] to
believe that we had actually made an
image by bending X-rays, and they tended
to regard the whole thing as a big fish
story."
Nobody believed that this was a viable
way forward. [music] And unfortunately,
the reaction was at least somewhat
justified.
First, this light isn't naturally
produced by anything on Earth. The
closest natural source is the sun.
>> We had to basically build an artificial
sun here on Earth.
>> Most scientists, including Kineshida,
produced X-ray light using a particle
accelerator or a synretron.
>> It gives an enormous amount of power.
It's as big as a soccer field. You can
fuel a whole FAB. The problem is if the
light goes out, the whole FEB goes out.
>> So each machine needed its own power
source. But even if you could produce
the light, you'd need to make incredibly
smooth mirrors to actually focus and
print those tiny features. You would
need the smoothest objects in the
universe. Okay, so I got a football and
I've got a bouncy ball and a cobblestone
street. Now, what do you think is going
to happen when I drop them?
the football basically bounces straight
up, but for the bouncy ball, it just
shoots off to the side. And that's
because the surface is relatively flat
for the football, which is much larger,
but it's super rough for the bouncy
ball. And a similar thing happens with
mirrors. If the surface is super rough
compared to the size of the wavelength,
then the light scatters randomly. Now,
it might look smooth, but if you zoom
into a mirror, you find something that
looks like this. You find all these
crazy bumps. And now to measure the
roughness, what you do is you take the
average of these bumps and that will
give you your mean line. Now for a
normal household mirror, the average
height is about 4,000 silicon atoms.
But for Kinoida's mirrors, which not
only needed to reflect X-ray light,
which has 100 times shorter wavelength,
but also needed to minimize scattering,
you know, so that all the photons make
it onto the wafer, it needed to be way
more smooth. It needed to be atomically
smooth. In fact, the average bump could
only be about 2.3 silicon atoms thick.
>> If one mirror would be the size of
Germany, the biggest bump would be about
a millimeter high.
>> But Kinoida refused to give up.
>> However, my belief did not change.
>> And soon, help would come from an
unlikely place. Across the Pacific
around 70 km east of San Francisco, is
Lawrence Livermore National Lab. a lab
that was born out of the cold war,
heavily funded by the US government and
built for one purpose and one purpose
only, nuclear weapons. The lab was
founded by the inventor of the cyclron,
Ernest Lawrence, and the father of the
hydrogen bomb, Edward Teller. And over
its lifetime, they designed over 10
fusion type nuclear warheads. So part of
their research focused on what happens
inside nuclear fusion reactions. fusion
reactions released a lot of X-ray light.
Light that they had never been able to
capture and analyze. But now using those
special multi-layer mirrors, there was a
chance. One of the scientists tasked
with making this work was Andrew Hover.
And within a few years, he and his team
used multi-layer mirrors to reflect some
X-ray light. But then in 1987, Andy got
a visit from a professor from Cornell.
He was very impressed with the
technologies that we developed and he
looked at me at the end of the day and
said this is all very interesting and
very neat and stuff but his m his words
and I'll remember it to the day I died
was can you do anything useful with this
stuff
this was the day before a Christmas
shutdown in 1987 and I was so inflamed
by that that comment that I went home
and for the next 10 days I wrote up a
multi-page white paper
>> he applied these mirrors to ligraphy to
print chips using X-rays. Around 5
months later, Andy presented his
findings at a conference. But like
Inoshida, it was not the response he was
hoping for.
>> It was extremely negative. That was the
low point of my career. I was literally
laughed off the stage and I kid you not.
every um person who I looked up to in
the field, they were listening to my
talk and they came up to the microphone
and told [music] me basically why it
wouldn't work, how stupid an idea it
was. Later that week, I flew back and
the following Monday, my boss asked me,
"How did it go?"
And I looked at him and I said, "I will
never speak of it again."
But then 3 days later, he gets [music] a
phone call from someone named Bill
Brinkman from Bellaps.
>> So I walked over to my boss and I said,
"Um, just got this phone call from a guy
named Bill Brinkman. Do you know who he
is?" And my boss's eyes popped open and
said, "Yeah, and he's the executive vice
president of AT&T." [music]
And I said, "Well, he just called me and
asked me to fly out to New Jersey and
give a talk." The response from my boss
said it all. Um, he basically said,
"Well, you got to go.
At Bell Labs, Andy found fellow
believers, and it couldn't have come at
a better time. Over the past 30 years,
the US government had invested billions
of dollars into national labs to
maintain the country's technological
edge during the Cold War. But by the
late 1980s, the Cold War was slowing
down, and all these labs were sitting on
research that had commercial potential.
So the government encouraged the labs to
partner with US companies to turn that
research into products and to stimulate
the economy [music] and the government
would then supply seed money. And so
Bell Labs partnered with Andy's labs and
two others to keep developing X-ray
litography.
And by 1993, the first international
conference for [music] X-ray litography
was held in Japan near Mount Fuji. In
the opening address, Kinoshida said that
as long as we do not lose the desire
that has sprung from within us,
technology will steadily advance from
the micro to the nano to the pico. They
even gave the technology a new name,
extreme ultraviolet lithography, or just
EUV.
But then in 1996, the US government cut
funding for the project. This spelled
disaster for the big chip companies like
Intel. The industry estimated that the
193 nanometer lithography tools would
fall behind Moore's law by 2005,
but there were no other alternatives.
So, Intel, Motorola, AMD, and other
companies got together and invested $250
million to keep it going, making it the
largest investment ever by private
industry in a Department of Energy
research project. By the year 2000, the
labs had produced this, the engineering
test stand. [music]
It was the first fully functioning EUV
prototype. It produced 9.8 watts of 13.4
nm EUV light which was then reflected by
8 mirrors from the source to the mask to
the wafer. It could print 70 nanmter
features and it proved that EUV could
work.
>> It was a milestone to get the
engineering test stand to work. It
demonstrated to people like Intel that,
you know, good engineering will get us
there. And then it seems like you've got
the prototype. Shouldn't be too hard to
then commercialize it.
>> That's what they thought. [laughter]
>> But the prototype had a major flaw. It
could only print about 10 wafers per
hour. And to make EUV economically
viable, it would have to print hundreds
of wafers per hour, 24/7, 365 days a
year. The main reason output was so slow
was because the light reflected off of
eight mirrors and the reticle, which is
also a mirror, just with the design
imprinted. Traditional masks that allow
light to pass through don't work
because, well, they absorb all the
light. Each mirror had a reflectivity of
around 70%, which is close to the max.
But after nine bounces, you're only left
with 4% of the light, which means that
out of every 100 photons, only four make
it to the waiver.
So, you might think just use way fewer
mirrors, but that only works up to a
point. When you focus light with any
optical system, you always get some
distortion. For example, rays that pass
through the outer edges of most lenses
focus light slightly different from
those near the center. This is called
spherical aberration. And normal cameras
correct for this and other aberrations
by using multiple lenses. And a mirror
system is no different. You need to have
a certain amount of mirrors before you
can say I have uh my aberrations under
control. In reality, the systems of
today have have six mirrors. That helps
a little, but after reflecting off six
mirrors and the reticle, you're still
only left with around 8% of your light.
So, they needed to drastically increase
the source power to at least 100 W. Now,
to most companies, that 10-fold increase
seemed impossible. Even people who
worked on the engineering test net noted
that while EUV technology itself is a
done deal, there were 6 zillion
engineering challenges to make it a fab
line reality. And so one by one,
American companies walked away from
developing a full UV litography machine.
That left just one company, ASML.
[music]
ASML, which used to stand for advanced
semiconductor materials litography, is
located in a small nondescript town in
the Netherlands. It spun off from
Phillips back in the 80s with little
more than a chat and a barely working
wafer stepper to its name. But
Philillips also gave them people. Yoss
Benrop ASML's first researcher and
Martin Fonden Brink who would eventually
become ASML's CTO and EUV's greatest
champion
>> and he's really like the Steve Jobs of
lithography and he saw EUV coming.
>> ASML had joined the US EUV consortium
earlier and now it became their task to
find a way to commercialize EUV. They
would work together with their German
partner Zeiss where Zeiss would take
care of the mirrors and ASML would focus
on the light source. One of the first
decisions when making any lithography
system is deciding which wavelength to
use.
>> In the early days, anything between 5
and 14 nmters was was explored.
>> Okay.
>> The the thing is you need to find a
source and you need to find optics that
reflect the wavelengths, right?
>> So, you have to look for the
combination. Underwood and Barbie had
already made mirrors that could reflect
light of around 4 nmters. And since that
wavelength is so small, it seems like
the obvious choice, but the maximum
reflectivity for those mirrors was only
around 20%. So after hitting six
mirrors, and theoretical, you're just
left with 0.00128%
[music] of the light, which is way too
low. Fortunately, further researchers
also looked at two other pairs, silicon
and malipinum, which had a theoretical
maximum reflectivity of 70% for
wavelengths around 30 nanometers and
malipinum and burillium with a
theoretical maximum reflectivity of 80%
for wavelengths around 11 nmters. So,
the choice seemed obvious, right? I
mean, pick the shorter wavelength and
the higher reflectivity. But it turns
out that burillium is extremely toxic
and it's also difficult to handle. So
scientists focused on silicon and
malipinum instead. To make the mirrors,
Zeiss used a process called sputtering.
A target of coating material is
bombarded with either plasma or ions
causing atoms to be ejected, fly off,
and stick to the mirror. This is a messy
process. So the layers end up having
bumps and gaps.
It was a nice trick that actually uh the
team in the Netherlands perfected with
iron beam. You just shake it a little
bit until the atoms falls in the hole
where it needs to be and then it's all
flat.
>> With the mirror design locked in, ASML
needed a source for that specific
wavelength.
>> So it was 13.x.
>> Yeah.
>> Now the next good question is what's the
X?
>> Now you look for the Now you look for
the source. So there are basically three
ways to generate EV uh to build an
a sun on Earth.
>> The first method which early researchers
used was the synretron, but it was
quickly rolled out because each machine
needed its own source. The other two
methods are based on the same principle.
When an electron recombines with an ion,
the ion drops to a lower energy level
and it releases that excess energy as a
photon. And if you choose the ion just
right, then that photon will have
exactly the wavelength you need. Now,
there are two ways you can create those
ions. The first is you take a metal,
heat it up until you get a metal vapor,
and then you apply a strong electric
field across it. This causes free
electrons to knock into nearby atoms
[music] and ionize them. If you then
turn off the electric field, the
electrons recombine with the ions and
produce light. This is discharge
produced plasma.
>> That's the concept we use first
>> because of its relative simplicity. And
we quickly got [clears throat] it to a
few watts. We wanted to get 100 W and we
struggled forever.
>> So you couldn't scale it.
>> We could not scale it.
>> They needed a drastic change. So they
switched to the second method. This
method uses a high-powered laser to hit
a target material, creating a plasma
that's more than 220,000°
C hot. The electrons have so much energy
that the nucleus can't hold on to them
anymore, and up to 14 electrons escape
their orbits. After the laser shuts off,
the electrons and ions recombine to
produce light. This is laser produced
plasma, and it was the only method that
seemed scalable.
In fact, this was the same method that
the engineering test used, a 1700 W
laser fired into a stream of gas to
produce 13.4 nanometer light. Buton had
a big problem. The conversion
efficiency, that is the ratio of usable
light to the amount of power you put in,
was terrible. It was only around 0.5%.
That's because while xenon does emit
light in the 13 to 14 nm range, there's
much more light released around 11 nm.
So, most of the energy went into making
light that the mirrors couldn't reflect.
Plus, the laser didn't ionize all the
atoms. So, leftover neutral xenon atoms
would strongly reabsorb some of that
13.4 nanome light.
So ASML started looking at another
material tin. Now tin has a much higher
emission peak around 13.5 nanometers
which results in a 5 to 10 times higher
conversion efficiency than xenon. But
just like neutral tin atoms also absorb
EUV light. So they came up with a crazy
idea to shoot one tiny tin droplet at a
time. But to get the required power you
would have to make and hit thousands of
droplets every second. all of which have
to be the exact same shape and size.
But it turns out that you can't
instantly make thousands of tin droplets
that are the exact same. So they found a
workaround. To make the droplets,
extremely pure tin is melted and pushed
through a microscopic nozzle by high
pressure nitrogen. This nozzle vibrates
at a high frequency, breaking the stream
into tiny droplets. These droplets are
irregular in size, shape, velocity, and
distance. And the whole process is
[music] chaotic.
>> That's like our magic sauce is how do
you modulate that tin jet so it forms
the droplets we want and that they're
stable.
>> I think we found some paper uh that
described this process. And it was sort
of eyeopening to me that it seems like
all the droplets actually come out
irregular [music] out of the nozzle, but
then before they reach the side where
they get hit by the laser, like the
little irregular droplets come together
to form these perfectly spaced,
perfectly regular droplets that are
about the same size and [music] shape
and all traveling at the same velocity.
That feels like magic to me, [music]
Jason.
>> Yeah, it's it's exactly that. is how do
you take a long stream of a tinjet
that wants to break up into all these uh
irregular droplets and like force onto
it that it's going to collapse into a
single droplet and then happen again and
again and again.
>> You also don't have that many variables
to play with. You've got the pressure
with which you push out the tin and the
frequency of the nozzle. Yeah, it seems
like a hard problem to solve. there's
not a whole lot of variables to play
with and so mastering that modulation of
the jet is is how we make the droplets.
But these droplets not only have to be
identical, they have to be moving
incredibly fast.
>> What will happen is if the next droplet
that's coming down the line is too
close, then it'll actually get like
disturbed and mess up the next plasma
event. So we have a requirement which is
both that we make 50,000 droplets per
second but also that they're traveling
extremely fast.
>> By 2011 their laser produced plasma
source reached 11 watt which was more
than double what they managed with their
previous source. But they were still
limited to just five waivers per hour.
So they needed to increase the power and
fast because they promised they'd hit 60
waivers per hour by the end of 2011.
Unfortunately, this new method had a
major flaw.
>> Now, the problem with the tin issue, you
hit the droplet, you generate EUV with a
very decent conversion efficiency. Where
does the tin go? Because like uh you
know 30 cm away, you have this
atomically flat, very beautiful, very
expensive mirror from our friends at
Zeiss.
>> Yeah.
>> And in the early days, we would coat the
thing within like this.
>> These machines need to run for a year.
You're putting lers of 10
through this plasma event and a single
nanometer of 10, if it was to land on
that collector mirror, you'd have to
take the collector out of commission. We
need to keep it almost perfectly clean
for for a year. [music]
>> Yeah. How do you even approach that?
>> So, our our our main tool here is the
hydrogen gas actually.
>> They filled the chamber with low
pressure hydrogen. This slows and cools
the tin particles down. And even if some
tin makes it to the collector, the
hydrogen pulls it off to form a gas
called stenain. This way, the machine
cleans the collectors while it's
running. But that hydrogen gas also gets
hot from all [music] those tin
explosions. So they need to keep
flushing new, cooler hydrogen into the
system while flushing out the stenane
and hotter gas. But they had to get the
pressure and the flow rate just right.
[music] I mean, too little hydrogen and
the mirrors would get too dirty, but too
much hydrogen would not only absorb too
much EUV light, but it would also cause
the system to overheat. But the question
is, how much heat is there? How much
energy is being deposited into the gas
and we were stumped for quite some time.
If you look at a EV light source, what
you'll see is that it's it's kind of
like a globe of like purplish red light
and you kind of ask yourself like why is
that happening? So, we bought an
ultraast camera. What we realized is
that after every plasma event, there's a
shock wave that goes propagating out
into the hydrogen gas. And it's
extremely repeatable. And you think to
yourself, there must be like an
explanation for this. And there's this
formula, the Taylor vonov formula that
explains point source explosions in an
environment from like say a nuclear
blast out to like supernova. So, I took
this formula and it like exactly
describes the data. It's just fantastic
that we're seeing these like little tiny
little supernovas happening in our
vessel 50,000 times a second.
>> And is that a fair way to think about
this? Like creating mini supernova.
>> Yeah, it's actually pretty similar. It's
almost like very similar to like a type
1A supernova. It turns out where you
kind of have an object that just fully
evaporates and explodes apart. And when
all that energy goes into the hydrogen
gas, it produces a a shock wave, a blast
wave that comes flying out, which is
basically the same thing. If you look up
in the night sky, there are these like
remnant supernovas that you can see
coming from space. Using those energy
calculations, they discovered they
needed to flush the hydrogen at
incredibly high speeds, around 360 km/h.
That's more than a category 5 hurricane,
even if you know those speeds are at low
density.
But 2012 came and went and they still
didn't have enough power. In fact, by
2013, ASML just reached 50 watts by
shooting 50,000 tin [music] droplets per
second. But this increased power came at
a price because more power means more
heat. Heat that ends up slightly
shifting the mirrors, resulting in
misaligned light and misaligned chip
layers. So, SAI built a nervous system
directly into the optics. robot guided
sensors that constantly measure the
exact position and angle of each mirror
down to the nanometer at the pico radian
which is [music] absolutely insane.
>> So how accurate do we need to control
this mirror? Now one of the things you
can do a thought experiment and I can
place a little laser on the side of this
mirror. Then we go all the way to the
moon and we put a dime here. So then
this light travels all the way here and
then with the accuracy I can control
this mirror.
>> Yes,
>> I can decide whether I point to this
[music] side of the dime or whether I
point to this side of the
>> What?
>> That's crazy.
>> So you can see that the pointing
accuracy [music] is uh that's also in in
pico radiance. Uh that is something very
extreme.
This allowed them to control the light
even when the power increased. While
size was doing a stellar job with the
optics, ASML was still struggling with
the power source. The problem was that
the tin droplets were too dense, meaning
that most of the emitted EUV light was
still getting reabsorbed by the neutral
atoms before it could ever reach the
collector [music] mirror.
>> The way we blasted the droplet was so
not enough light, too much debris. To
make matters worse, they could see that
about 10 years from now, they would need
a new generation of machine, a high NA
EUV machine. Essentially, one with a
larger optic system that could print
smaller features. So, what did they do?
They decided to double down and invest
in the next generation before they even
got the current one to work.
>> The most doubtful [music] period was in
the beginning. So, I started to work on
this in 2012. By that time, EUV was not
working and there was this uh crazy
idiot working on the next generation
where we could not even make the EUV
light in the first place.
>> Not only are you all in on EUV, you're
doubling down even before you know if
EUV is going to work.
>> Yes.
>> But to keep funding the development,
they needed money and lots of it. So,
they turned to the very people who
needed this technology.
>> ASML reached out to its main customers.
Okay, you want this technology for the
next generation of chips? Well, you need
to make us able to invest more by
investing in us.
>> Intel invested around $4.1 billion and
Samsung and TSMC invested another 1.3
billion combined. So, they can keep the
research going. But with no product to
show, customers were running out of
patience. We were crucified at every
conference that the promises we made
last year we we were unable to live up
to.
>> Yeah.
>> And he said, "This is what you showed
two years ago. This is what you show
last year. This is what you're telling
me this year. So why would I believe
you?"
>> They were getting desperate.
>> But this was I think about 2012 or or
13. We were struggling to get the EV
power up and Kinoshida visited us. I
took him to dinner in a small town
nearby and across from the restaurant
was a Maria chapel and now you know
science we have come to the limits of
science hey let's go for define
intervention so we went to the chapel so
Kinoshita just to be safe lit three
candles for the three suppliers that
were pursuing EV technology at the time
and lo and behold and I have the data to
prove it there is a very strong
correlation between us lighting the
candle and uh power going up. [laughter]
It's not a causal effect, but there is a
strong correlation.
>> The big idea was instead of hitting the
droplet once, hit it twice.
>> One shot to hit the droplet and it
expands in like a pancake shape.
>> Yeah.
>> And then only then have the second shot,
the more powerful main pulse where you
evaporate the pancake and turn it into a
plasma. This was a major breakthrough.
>> By changing the target from a droplet to
a pancake, you got a larger surface area
for the laser to vaporize, but without
the cost of adding more debris or
neutral atoms, because now the tin is
vaporized all at once. By 2014, they
finally managed to hit that coveted 100
W mark. But improvements in
multi-patterning with 193 nanometers now
meant that EUV would only be useful if
the source reached [music] at least 200
W and made 125 waiverss per hour.
>> The source went from 100 to 200. But as
the industry moved on, nobody waits for
you, you know, they find other
solutions.
>> We had to catch up. So it was a moving
goalpost.
>> One of the problems was how do you
perfectly time the laser so you hit each
of these droplets? So the analogy is a
bit like a golf ball that you need to
land in the hole 200 m away. Not like
land on the green, not bounce and then
get in the hole, but like land in the
hole every time. That's the level of
precision that we need to deliver the
droplets. Those droplets are traveling
through this like mastrom of hydrogen
flow. The speeds are tremendously high.
It's like shoot golf balls through a
tornado and then right when it lands at
the hole, that's when it needs to get
hit by the laser. So in order to
basically track the droplets for that we
use laser curtains and we can sort of
look at when does the droplet pass
through a laser curtain. Those scattered
photons tell us basically when and where
is the droplet and then importantly
tells us when to fire the laser. So we
actually have to take into account how
long will it take for the light pulse to
hit the droplet after we send the pulse.
Now, by 2015, they were getting closer
and closer to that coveted 200 W mark
when all of a sudden the ASML board
members got summoned. Th this was one of
these decisive moments where our
customers were really thin on patience.
Martin and all the board members were
summoned to Korea to show 200 W and they
were really fed up with it, you know.
Yeah, you either show it now or you you
>> go away. And when they entered the
plane, the experiment was still running.
>> Okay.
>> When they exited the plane, they had the
first result demonstrating to this is
how close we came.
>> With the source power up, there was one
final problem that had to be solved
before they could begin manufacturing
their machine. See, while the hydrogen
gas did protect the collector mirror
from debris, it wasn't perfect. All the
intense high energy photons and hydrogen
ions zipping around deteriorated a very
special top coating on the collector. So
they still had to clean the mirrors
every 10 hours which you know is
terrible for productivity. Martin
Funbrink asked for updates every day on
their progress. But then one of the
engineers noticed that every time they
opened up the machine the mirrors
suddenly seemed cleaner.
>> That he kind of chimed in and said, "Oh
uh wait a second. Whenever we open up
the machine, oxygen comes in and our
problem is solved. Couldn't we think of
a way to add just a little oxygen to our
system and make sure that the collector
stays clean longer? And so they started
experimenting uh with the amount of
oxygen uh that was needed in the vacuum.
And they finally got to this point,
okay, if we add so much oxygen, we'll
keep the collector clean for longer.
>> With this fix, ASML's machine could run
continuously for much longer. and it
finally became commercially viable. By
2016, orders started pouring in and now
all of the most advanced chips [music]
need ASML's machine, making them perhaps
the most important tech company in the
world. ASML's first commercial machines
had a numerical aperture of 0.33 and
could print 13 nanometer lines. These
are called the lowa machines and ASML
still makes them. But the machine that
Jon's team started working on back in
2012 was the next generation which had a
larger [clears throat] optic system so
they could print even smaller features.
This is the high NA machine with a
numerical aperture of 0.55. [music]
And we get to see their latest version
up close.
>> How much is the machine?
>> Uh we always say north of €350 million.
>> And you can actually buy it, right?
>> You can if you want. Yeah.
>> If I had the money, I could buy it.
>> Yes, you could.
How many people have seen this before?
>> We really limit the amount of people
that get to go inside the clean room.
>> ASML's machines are built in a super
strict clean room. In any cubic meter,
there can be no more than 10 [music]
particles, only 0.1 microns large, and
nothing bigger than that. A spec of
Poland is around 20 microns, and
extremely fine sand is around [music] 10
microns. To put all of this in
perspective, hospital operating rooms,
which have to be extremely [music]
clean, only allow a maximum of 10,000
particles per cubic meter that are 0.1
microns [music] wide.
>> It's so unfair how much better Mark
looks though. [laughter] It is light
suit. I feel like a little smurf.
>> Okay, so we're going to go through the
air showers. So, you're going to have to
do as I do.
>> Okay, so this is brushing down all the
particles that are still on.
>> Yeah. So, this is like super clean air
blowing us clean.
>> This place is huge.
>> It's huge.
>> It's insane. I've been in a clean room a
couple times before, but it's nothing
compared to this. Are there any secret
areas here where almost no one has
access to?
>> Uh, I can't tell you.
>> Great answer. Okay, so this is the total
system.
>> This is it.
>> This is crazy. Look how big it is.
This is the most advanced machine
humanity's ever built. It's [music]
taken many, many years, decades of
development, many billions of dollars,
all to get this humongous beauty. So,
this is the first highame machine. Yes.
So, if you saw pictures on the internet
or whatever,
>> Yeah.
>> that's this machine. So, the very first
lines ever printed at 8 nanometers and
stuff, that was this machine. The
smoothest object on Earth. Yeah. It's
all in here. Yeah.
>> Wait. So, let me see if I can figure
this out.
>> This is the light source. That's where
they make the extreme ultraviolet.
>> Yes.
>> And then the laser must come in from
Yeah. Let's take a look at the laser.
>> In fact, we got to see just how the
laser and light source work.
>> I think we're entering the laser system
here. Mark's just making sure, I think,
that we can actually film here. We're
not catching anything we're not supposed
to. Oh, wow. This looks dangerous. Now,
the laser system is covered by all of
these brown cabinets, but here is a
model version. A carbon dioxide laser of
just a few watts enters this amplifier
where it bounces around until it's
roughly five times its original power.
It then goes through a total of four
different amplifiers [music]
to bring the final laser up to 20,000 W,
which is four times stronger than lasers
that cut through steel. Over here we
have the the amplifiers that generate
this this powerful laser beam. Yeah.
>> And then it basically comes out and this
is part of the beam transport system.
>> Yeah.
>> Where it's brought to the machine.
>> So this pipe here has the big laser beam
and this this has a mirror.
>> Yes.
>> Then the pulses travel to the light
source module.
>> It kind of looks like a transformer or
like a I don't know like a spaceship.
There's so many wires going everywhere.
Don't touch this.
Holy crap. This is pretty big, huh?
>> This is insane. And this is just a light
source.
>> This is just a light source. Are you
getting this comparison shot? And so you
need all of this
>> just to make EUV light.
>> Just to make the light. That's
incredible. Can we do a little walk
around?
>> We can do a little walk.
>> Let's go.
>> So, basically, this is the heart of the
source.
>> Can I stand on here? Uh, if you're below
137, you can
>> I don't I think I am.
Woo. And so the thin droplets are coming
in from the left.
>> Yes.
>> Then we're shooting the laser from here.
>> Yeah.
>> Okay. It explodes
>> and then the light
>> the light goes out there.
One improvement from ASML's first EUV
machine to their newest one is the
number of pulses that hit the droplet.
The first prepulse still flattens the
droplet into a pancake. But now there's
also a second pre-pulse that further
reduces the density. It [music]
basically turns it into a low density
gas. It verarifies it. And then the
final pulse essentially ionizes all
[music] of it. So for basically the same
power coming from the drive laser, they
get even more EUV light. Now if they
want even more light, then the only way
to do that is by hitting more droplets.
[music]
And that's exactly what they did. our
most recent EV light sources that we're
shipping right now, which are around the
500 watt level. Uh we increased the
[music] rep rate up to 60,000 times per
second. And then we have a road map
that's going to go to 100,000 droplets
per second. We've actually now already
demonstrated this 100,000 droplets per
second in the lab. So it's not an if but
a win.
>> Crazy.
>> The three pulses that we use to make the
pancake, to blow up the pancake a little
bit, and then to evaporate the pancake.
Yeah. The first two pulses, they would
be coming in through this pipe here.
>> Yeah.
>> And then the main pulse with the big
laser, the laser beam would be delivered
through this pipe here. Both the high
and low machine shipping out right now
use three pulses and eventually they
will hit more droplets per second. But
the light source is just one small part
of the full machine. After bouncing off
the collector mirror, the EUV light
moves to the illuminator. A set of
mirrors shape and focus the light before
it [music] hits the reticle. The reticle
is the top half, and this module is
built in a separate facility and
installed later. Next, the light goes
into the projection optics box, which is
a set of mirrors that shrink the light
down. The high machine can shrink the
pattern eight times in the vertical
direction and four times in the
horizontal direction. The mirrors are
also much smoother still. If the low
NA's mirrors were the size of Germany,
the tallest bump would be about a
millimeter. But if the high mirrors were
the size of the world, the tallest bump
would be about the thickness of a
playing card. By the combination of both
of these improvements, ASML was able to
increase the numerical aperture from
0.33 to 0.55.
And finally, the light hits the waiver.
In order to print around 185 waivers per
hour, the reticle whips back and forth
at accelerations of over 20 gs. That's
over five times the acceleration of a
Formula 1 car. And this is some actual
footage of what that's like inside this
machine. And notice that this is not
sped up.
But the crazy thing to me about this
machine isn't how fast the reticle moves
or even how small it can print, but it's
just how insanely accurate it needs to
be. The most any two layers can be off,
which is called the overlay, is [music]
1 nanometer. That's five freaking
silicon atoms of precision. That's
insane. So typically what we do as
system engineers is that we make a
budget. [music] So we say hey you get
let's say a nanometer and uh and we
divide then [music]
uh the nanometers uh to to smaller
fractions
>> the nanometers total. It's not like you
your group gets
>> you get a nanometer you get no no you
get a nanometer in total. Yes. [music]
So you have to uh to fight for the for
your part of the nanometer.
>> [music]
>> It's kind of cool to realize that like
every smartphone nowadays has has a chip
that is made with a machine or that was
actually put together here.
>> So that's a cool thought.
>> Take a look at this.
>> It's pretty massive.
>> So big.
>> So do you cover it up?
>> Yes. At a customer fab it will be
looking like a big white box.
>> I like it better like this.
>> Yeah, me too.
[music]
>> It's funny. You need such a big machine,
so much infrastructure to make the
tiniest things
>> we can make at scale.
>> It's inversely proportional. Yeah.
Smaller you want to go, the larger
everything around it becomes. After the
machines are assembled, tested, and
approved, they are disassembled to ship
all around the world. 5,000 companies
supply 100,000 parts, 3,000 cables,
40,000 bolts, and 2 km of hosing. ASML
ships their hyena machine in 250
containers spread out over 25 trucks and
seven Boeing 747s.
Despite all the doubt and setback, EUV
finally made it to manufacturing level
three decades after Kinoida's first
im