Drawing smaller than light
Strip away the marketing and a chip is a stack of patterns, printed one layer at a time with light. The process is called photolithography, and the idea is the same one behind a slide projector: shine light through a stencil (a mask), focus the image down through a lens, and expose a light-sensitive coating on the wafer. Wherever the light lands, the coating changes, and that pattern becomes a layer of transistors or metal wires. Every other step in chipmaking exists to support this one.
Here is the problem that has shaped the entire industry: you are trying to draw features far smaller than the wavelength of the light you are drawing with. A modern feature can be a handful of nanometres across, while the light used for years was 193 nm — the feature is *tens of times smaller* than the very tool printing it. That is like trying to sign your name with a paintbrush wider than the signature. It sounds impossible, and physics agrees it should be hard.
The 193 nm wall
How small a feature you can resolve is captured by one famous relationship: the smallest pitch scales as k1 × λ / NA — the wavelength λ divided by the numerical aperture NA of the lens, times a process factor k1 that you fight to shrink. To draw smaller, you have three levers: use shorter-wavelength light (smaller λ), build a lens that gathers light from a wider angle (bigger NA), or get cleverer about the process (smaller k1). For a long stretch the industry had pushed all three to their limits — and was stuck at one wavelength: 193 nm, from an argon-fluoride laser.
The last big win on the NA lever was a beautiful hack called immersion: flood the gap between the lens and the wafer with purified water instead of air. Light bends more in water, which effectively raises NA and squeezes the printable feature smaller — without changing the wavelength at all. 193 nm immersion lithography became the workhorse of the leading edge for years. But water was the end of that road. There was no obvious way to shorten 193 nm further, and immersion had wrung out the NA. The industry had hit a wall — and chips were supposed to keep shrinking anyway.
Multi-patterning: brute force
When you cannot make the light finer, you cheat with arithmetic. Multi-patterning is the brute-force answer: if a single exposure cannot resolve features packed tightly enough, split that one layer into several coarser exposures and interleave the results. Print every other line in one pass, then shift the mask and print the lines in between on a second pass. Two masks now do the work of one — and the final pattern is twice as dense as either mask could manage alone.
It works, and it kept Moore's Law alive through the 193 nm era — but the cost is steep, and it compounds. Each split means another mask (masks are expensive), another full exposure, and extra deposition and etch steps to transfer and align each pass. Worse, the passes must line up to within a fraction of the feature size; any misalignment between them — called overlay error — directly corrupts the layer. By the densest layers, a single layer might need two, three, or four separate patterning passes (LELE, SADP, SAQP, in the jargon). More masks, more steps, more cost, more ways to go wrong. Multi-patterning is the sound of a tool being pushed well past its comfort zone.
FEATURE PITCH vs. WHAT IT TAKES 193 nm immersion, single pass light: ~~~~~ 193 nm wavelength ~~~~~ prints: | | | | | (coarse — 1 mask) 193 nm immersion + multi-patterning mask A: | | | mask B: | | | (+ extra mask, etch, align) result: | | | | | | | | | | (dense — 2-4 masks per layer) EUV, single pass light: . 13.5 nm wavelength . prints: |||||||||||||||||||||| (dense — 1 mask)
EUV: 13.5 nm light
The real cure is the obvious one the equation pointed at all along: shorten the wavelength. Extreme ultraviolet lithography does exactly that, and not by a little — it jumps from 193 nm all the way down to 13.5 nm, more than a tenfold reduction. At a stroke, the same density that needed three or four masks under 193 nm immersion can be printed in a single EUV exposure. The wall the industry spent a decade clawing over with multi-patterning is, for the most critical layers, simply gone.
The catch is that 13.5 nm light is monstrously difficult to make and to handle, because at that wavelength *everything* absorbs it — including air, and including glass. That single fact forces three extraordinary engineering choices. First, the whole light path runs in vacuum, because even a few centimetres of air would swallow the beam. Second, you cannot use lenses; the optics are all mirrors — exquisitely flat, multilayer-coated reflectors, since no transparent material exists at this wavelength. Third, the light itself is born violently: tiny droplets of molten tin are zapped twice each by a high-power laser, flashing into a plasma that radiates at 13.5 nm. Tens of thousands of droplets per second, each hit on the fly.
High-NA EUV
EUV bought a huge jump in wavelength, but Moore's Law keeps asking for more. With λ now fixed at 13.5 nm, attention swings back to the other lever in the equation: NA, the numerical aperture. High-NA EUV is the next generation of machine, raising NA from 0.33 (the first EUV tools) up to 0.55. A bigger NA gathers light over a wider angle, which sharpens the smallest printable feature and pushes single-exposure patterning to the next node or two — delaying the day EUV itself needs multi-patterning.
Nothing is free. To bend EUV light through that wider angle, the new optics are *anamorphic* — they magnify differently in the two directions — and the practical consequence is a smaller exposure field, roughly half the area printed per shot compared to a 0.33 NA tool. A large chip that used to fit under one exposure may now have to be split across two and stitched back together. So the high-NA story is the same trade you have seen all the way down this rung: a finer pen, bought with a smaller page and an even more expensive, even rarer machine.
Why litho dominates cost
Step back and the pattern is unmistakable: lithography is the gravitational centre of leading-edge cost and complexity. The scanners are the most expensive tools in the fab by a wide margin. The masks are costly, and multi-patterning multiplies how many you need per layer. Every exposure drags a train of deposition, etch, and alignment steps behind it, so the number of patterning passes is one of the biggest knobs on the total cost — and on yield, since every extra step is another chance for a defect.
This is why so much of modern chip strategy quietly orbits lithography. The eye-watering price of building an integrated circuit on a cutting-edge semiconductor process is, in large part, the price of patterning it — and that single fact is the unspoken reason behind much of the frontier. It pushes designers to print fewer hard layers, to co-optimize the design and the process together, and (as later guides show) to break a big chip into smaller pieces so that only the parts that truly need the most expensive lithography pay for it. Patterning at the limit is where the leading edge is won, lost, and paid for.