Every transistor on a leading-edge chip is printed with light so fine it stops behaving like a wave and starts behaving like a hail of individual particles. At 13.5 nanometres, a single printed feature is exposed by a countable number of photons — and because that count is random, the edge of the feature is random too. That randomness, not any flaw in the machine, is one of the primary causes of defects at the 3 nm and 5 nm nodes.
Light you can count
Energy and wavelength are locked together: a photon of wavelength λ carries energy E = hc / λ At the EUV wavelength of 13.5 nm that works out to about 91.8 electron-volts per photon — roughly fourteen times the energy of the 193 nm deep-UV light that printed chips a generation ago.
Because each photon carries a fixed, relatively large amount of energy, a given exposure dose corresponds to a specific — and surprisingly small — number of photons landing per unit area. At a typical 35 mJ/cm² dose, only about 24 photons arrive per square nanometre. A tight contact hole a few nanometres across is therefore defined by a few hundred photons, or fewer. You are, quite literally, printing circuits by counting a modest handful of light particles per feature.
Poisson statistics and the √N law
Photons arrive independently of one another, like raindrops on a paving stone. That makes the number landing in any small region follow a Poisson distribution, which has one defining property: its variance equals its mean. If you expect N photons, the standard deviation of what you actually get is
σ = √N → relative fluctuation = σ / N = 1 / √NThis single relationship explains the whole problem. At N = 10,000 photons the relative swing is just 1%. At N = 100 it is 10%. Shrink the feature or drop the dose and the relative noise grows as one over the square root of the count — so the smallest features at the lowest doses are exactly where the randomness bites hardest.
From a fluctuation to a defect
A photoresist develops where it has absorbed more than a threshold amount of energy, and stays put where it has absorbed less. When the local photon count happens to fall below that threshold — or a neighbouring region happens to exceed it — you get a feature that is missing, broken, merged, or misplaced. At billions of features per chip, a random failure rate of even one in a million becomes a yield problem.
This is why engineers talk about a stochastic floor: photon shot noise sets a hard lower bound on defectivity that no amount of tuning the optics can remove. You can only push it down by raising the photon count — more dose, larger features, or more efficient absorption — each of which trades against throughput, resolution, or resist design.
Halving a feature's area, or halving the dose, multiplies its relative shot noise by √2. There is no way around the √N law — only levers that change N.
Why EUV makes this sharper than DUV
Counter-intuitively, EUV's short wavelength works against photon statistics. Because each EUV photon carries ~14× the energy of a DUV photon, a given dose is delivered by roughly 14× fewer photons. Fewer photons per feature means larger relative fluctuations for the same energy on the wafer. EUV buys you resolution — the ability to draw features that small in the first place — at the cost of a grainier exposure, which is precisely the regime where stochastic defects appear.
EUV Photon Shot Noise Calculator
Enter a dose, wavelength, and feature size and watch the photon count and its 1/√N uncertainty update live.
What shot noise is not
Photon counting is the floor, not the whole story. The path from "how many photons landed" to "did the feature print" also runs through secondary-electron blur, resist chemistry, and a development threshold — each with its own statistics. And whether a given fluctuation actually becomes a defect is itself a probability, which is where the stochastic defect tail comes in. Predicting where on a real layout those defects will land, rather than just how likely they are on average, is the problem the SECD-CRF project is being built to solve.