Carved by a Light Too Fat to Carve It

TL;DR

• Light is a wave, and a wave cannot be focused to a spot much finer than its own wavelength. This is the diffraction limit, written down by Ernst Abbe in 1873; it is physics, not an engineering excuse.
• The light that prints chips is deep-ultraviolet at 193 nanometers, yet the features it draws are a fraction of that. The limit was never broken. It was routed around, one clause at a time.
• Move one: make the lens collect more of the light by raising its numerical aperture, then flood the gap with water to push that number above 1 (immersion).
• Move two: bend the wave before it lands, with phase-shift masks and tilted illumination, so the fine detail actually survives the trip through the lens.
• Move three: when one exposure cannot resolve a dense pattern, print it in two or four coarser passes (multiple patterning).
• Move four: change the pencil. Extreme-ultraviolet light at 13.5 nanometers is roughly fourteen times finer, but no lens on Earth can focus it, so it is bounced off mirrors in a vacuum instead.
• The takeaway: the diffraction limit constrains a single exposure of a given wavelength through a given medium. Change any one of those and the wall moves. "5 nm" is the name of a manufacturing generation, not the size of anything on the chip.

A sliver of silicon the size of a fingernail, inside the phone in your pocket, carries tens of billions of transistors: more switches than there are people on Earth. Each is far smaller than a speck of dust. Each is smaller, in fact, than a single wavelength of visible light. The graphics chip in a laptop or a game console is the same feat on a larger slab. Almost everything digital you touch works because we learned to draw at this scale, reliably, billions of features at a time.

We do it by printing. A chip is not carved one transistor at a time; the whole circuit pattern is exposed at once, photographically. Light shines through a patterned mask, the way a stencil shapes spray paint, onto a wafer coated in a light-sensitive film, and where the light lands the film changes so the pattern can be etched into the silicon beneath. That is lithography, literally "stone-writing," and doing it in a single flash is what lets one chip carry billions of identical parts. The whole trick lives in that light.

And here is the trouble with light. Hold a fat paintbrush over paper and try to paint a hairline: you cannot, because the paint spreads to the width of the bristles no matter how lightly you press. The finest stroke is set by the tool, not your intention. Light is that brush. Focus a beam to the smallest possible spot and you get not a point but a small, soft disk, its width set by the wavelength. A wave refuses to be squeezed into a region much smaller than itself.

Now the puzzle. The light that prints today's chips has a wavelength of 193 nanometers, yet the features it prints are smaller than that wavelength, in places by a wide margin: a brush drawing a line finer than its own bristles. The law that forbids this is real, and it was never broken. It was read very carefully and routed around, clause by clause.

Why can't you just use a finer pencil?

Because the pencil is a wave, and waves diffract: they bend and spread when they pass an edge or through an opening. That spreading is what blurs the focused spot, and it sets a hard floor on how fine an image any lens can form.

Ernst Abbe worked out that floor in 1873, building microscopes in Jena: the smallest separation a lens can resolve is roughly the wavelength λ divided by twice the numerical aperture, written d = λ / (2 × NA). NA = n × sin θ measures how wide a cone of light the lens gathers, where θ is the cone's half-angle and n is the refractive index of the medium in front of it. Lord Rayleigh reached the same criterion for telescopes a few years later. On a real immersion scanner, the smallest printable line shrinks in lock-step with λ / NA, exactly as the formula says it must.

Chipmakers write the same physics in their own units. The smallest printable feature, the critical dimension, is

CD = k₁ × (λ / NA)

Here k₁ (say "k-one") is a single number that bundles up every optical trick the engineers know. For a single exposure it cannot fall much below about 0.25, the hard limit of two-beam interference; a real production process runs a little above that, near 0.29. So the limit has exactly three handles: make λ smaller, make NA bigger, or push k₁ toward 0.25. Every advance in the last forty years is one of those three, and we will take them in order.

Start with the wavelength already in the factories. For two decades the workhorse has been an argon-fluoride excimer laser emitting deep-ultraviolet light at 193 nanometers; the previous generation used krypton fluoride at 248. Plug in 193 nanometers, the best dry-lens NA of about 0.93, and the k₁ = 0.25 floor. The smallest line comes out near 52 nanometers. That is already smaller than the wavelength, which surprises people, but it is nowhere near a modern chip. We need the other handles.

How do you make the lens collect more light?

Look again at NA = n × sin θ. The angle term sin θ can never exceed 1, and real lenses top out around 0.93, so in air the numerical aperture is stuck below 1. That looks like a wall. But notice the other factor: n, the refractive index of the medium the light passes through just before it reaches the wafer. In air, n is essentially 1, so that factor sat there contributing nothing, a free multiplier no one was using.

Fill the gap with something denser than air and n goes up, and NA goes up with it. Put a thin layer of ultrapure water between the final lens and the wafer, and at 193 nanometers water has a refractive index of about 1.44. The numerical aperture, formerly capped below 1, now reaches about 1.35. This is immersion lithography, developed in the early 2000s and built into volume production by ASML in the Netherlands and Nikon in Japan. The same laser, the same wavelength, but the lens now drinks in a wider cone of light, and the printable feature drops well below the dry-system limit. Nothing about the light changed. We just stopped wasting the medium.

How do you bend a wave into a shape it does not want to make?

The third handle, k₁, is the strange one. The other two change the physics outside the wafer; pushing k₁ toward 0.25 means reshaping the wave itself on its way in, so that diffraction lands where you want it instead of where it would naturally go. Three moves do most of the work, and they share one idea: the pattern on the mask (the stencil the light shines through) need not look like the pattern you want on the wafer. It only needs to produce that pattern after diffraction has had its way.

The first move is the phase-shift mask. In 1982, Marc Levenson and colleagues at IBM realized that two adjacent clear openings in a mask print as a single blur when they sit too close, because their spread-out light overlaps in the gap that should stay dark. Their fix: etch one of the openings slightly deeper so its light emerges exactly half a wavelength behind its neighbor's. The two waves now arrive in the dark gap perfectly out of step, crest meeting trough, and cancel. Destructive interference, recruited on purpose to carve a sharp dark line where diffraction wanted a smear.

The second move is off-axis illumination: instead of lighting the mask straight on, tilt the incoming light. The fine detail in a pattern lives in the steeply diffracted light, the rays bent hard to the side, and those are exactly the rays a straight-on setup loses off the edge of the lens. Tilt the source and you slide those crucial rays back into the lens's cone, recovering the detail they carry.

The third move is optical proximity correction. Run the blurry print through a simulation, see how diffraction will round the corners and pinch the line-ends, then pre-distort the mask to fight it: add little hammerheads at the ends, serifs at the corners, assist features too small to print on their own. A modern mask, seen under magnification, no longer resembles the circuit at all. It is a deliberately deformed pattern whose blur, and only whose blur, is the shape you wanted.

What if one exposure is not enough?

Even with every wave trick, a single exposure hits the k₁ = 0.25 floor and stops. So you stop trying to do it in one exposure.

Suppose you need lines packed twice as densely as one shot can resolve. Print every other line in the first exposure, where they sit far enough apart to be legal; etch them into a hard layer; then print the lines that go in between, in a second pass offset from the first. Two coarse, legal patterns interleave into one fine, illegal-looking one. This is multiple patterning, and the simplest version is called litho-etch-litho-etch.

A cleverer version skips the second mask. Print a pattern of ridges, coat the whole surface with a conformal film so each ridge grows a thin sidewall of equal thickness, then etch away the ridges and the film tops, leaving only the sidewall spacers. Because every ridge had two sides, you now have twice as many lines as you printed, at half the pitch, and their width is set by a deposited film thickness controllable to a single atomic layer rather than by the blurry optics. Do it again on the result and you quadruple. This is self-aligned double and quadruple patterning. The cost is real: more masks, more deposition and etch steps, and a punishing demand that every layer line up with the last to within a few nanometers. But it beats a floor that one exposure cannot.

What if you change the pencil after all?

Every move so far keeps the 193-nanometer light and fights its wavelength. The last move surrenders and gets a shorter one. Extreme-ultraviolet lithography uses light at 13.5 nanometers, roughly fourteen times shorter than 193. Drop λ by fourteen and the whole diffraction limit drops with it; suddenly single-digit-nanometer features come within reach in a single exposure.

The catch is severe: 13.5-nanometer light is absorbed by almost everything. It does not pass through air, so the whole optical path sits in a vacuum; it does not pass through glass either, so there are no lenses, only mirrors. But ordinary mirrors barely reflect EUV; a single surface throws almost all of it away. The fix is a Bragg mirror: forty to fifty alternating molybdenum-silicon layers, each a few nanometers thick, spaced so the faint reflection off every layer arrives in step with the rest and they add instead of cancel. It is the same constructive interference the phase-shift mask used, now stacked fifty deep to build one good mirror out of fifty poor ones. Even so, each mirror returns only part of the light, and across a dozen of them only a small fraction survives to the wafer.

Making the light is its own ordeal. Drops of molten tin are fired across the machine and zapped in flight by a high-power carbon-dioxide laser, often with a first pulse to flatten the droplet and a second to vaporize it into a plasma hot enough to emit at 13.5 nanometers. Tens of thousands of droplets a second, each hit twice, in vacuum, the light then bounced off a parade of multilayer mirrors and finally off a mask that is itself a mirror. An EUV scanner is the most complex machine ever mass-produced, built by a single company, ASML in the Netherlands, and run by the handful of fabs that can afford it, led by TSMC in Taiwan. The newest "high-NA" generation, which lifts the numerical aperture from 0.33 to 0.55 to print still finer, costs on the order of several hundred million dollars per tool.

So how small is "5 nanometers"?

Here is the quiet payoff. Putting these moves together, the industry labels its manufacturing generations "7 nm", "5 nm", "3 nm". It is natural to assume something on the chip is 5 nanometers wide. Almost nothing is. The names long ago detached from any physical dimension; they are marketing labels for a generation of process technology. The smallest features on a "5 nm" chip are real, and they are bigger than 5 nanometers; the number stopped meaning a width years ago. They are built not by a 5-nanometer pencil but by the stack of moves in this article, each one defeating a limit that looked absolute by changing the conditions the limit was stated under.

That is the real lesson, and it outlasts any particular node. The diffraction limit is true. Abbe was right in 1873, and no chip has ever violated him. What the chipmakers noticed is that his law has fine print: it bounds a single exposure, of a given wavelength, through a given medium, collected by a given cone. Immersion edited the medium. Phase-shift masks and tilted light edited the cone and the wave. Multiple patterning edited "single". EUV edited the wavelength. The wall never moved because someone pushed harder on it. It moved because someone read the label on it and changed what it was a wall against.

None of this is free, and none of it is endless. Each move buys resolution with cost or complexity: immersion stops at the refractive index of water; the k₁ floor of 0.25 never moves for a single shot; multiple patterning multiplies the masks and stacks one alignment error on the next; and EUV bought its shorter wavelength with the most complex machine ever mass-produced. Even 13.5-nanometer light obeys Abbe, and its own diffraction limit is already in view. The chipmakers have changed the brush, the medium, the angle, and the number of strokes; what they have never done is repeal the law itself. The line on the chip is thinner than the light that drew it. How much thinner it can still get is a question the industry has not yet answered.

And even that question is only half the story. Everything here is about printing a feature: can light draw a line this fine? It says nothing about whether the transistor then works. Shrink the switch itself to a few atoms across and the electrons stop obeying it; they quantum-tunnel, slipping straight through a barrier they classically should not be able to cross, so the gate leaks and the "off" state is never quite off. That is not an optical wall you slip past with cleverer light; it is the device physics itself pushing back, and it is why the newest chips change the transistor's shape, not just its size. Where that second wall really stands is a different question from this one, and a story for another day.