Yield, defect density, and binning: Visual walkthrough

Step 1 — What we are counting: dust on a rectangle

WHAT. Picture the whole wafer sprinkled with random killer specks, and zoom into a single die.

WHY. Before any math, we must be crystal-clear about the event we're chasing: "this one rectangle caught no specks." Everything else is machinery to compute that one probability.

PICTURE. In the figure, orange dots are defects scattered over the wafer. The teal rectangle is our die. Ask only one question: did any dot land inside teal?

Figure — Yield, defect density, and binning

Step 2 — Defect density $D_{0}$ : how dirty is the process?

WHAT. We compress "how dusty is the whole process" into a single number $D_{0}$ . If a region has area $A$ , the average count of defects it catches is

$λ = defects per cm^{2} D_{0} \times cm^{2} of one die A .$

Read the units: $\frac{defects}{cm ^{2}} \times cm^{2} = defects$ . So $λ$ is a pure expected count of defects on one die. We give it the name $λ$ (Greek "lambda") because it's the one number the whole derivation pivots on.

WHY multiply? Density times area is always "how much stuff is in this region" — the same way (people per km²) × (km²) = people. $D_{0}$ is a density; $A$ is the region; their product is the count.

PICTURE. The figure shows two dies of different sizes on the same dusty wafer. The bigger die overlaps more dots — so its $λ = D_{0} A$ is larger. Same $D_{0}$ , bigger $A$ , more expected defects.

Step 3 — Slice the die into many tiny boxes

WHAT. Cut the die into $n$ equal tiny squares, where $n$ is huge. Each box is so small it can hold at most one defect — it either has a speck or it doesn't.

WHY do this? A tiny yes/no box is something we can reason about with plain counting. We turn one hard question ("how many specks on the whole die?") into $n$ easy coin-flip questions ("does this box have a speck?"). This is the trick that will manufacture $e$ for us.

The chance one box is hit. The die averages $λ$ defects spread over $n$ boxes, so each box averages $λ / n$ defects. Since a box holds at most one, that average is its probability of being hit:

$p = \frac{λ}{n} (tiny, because n is huge) .$

PICTURE. The die is now a grid of $n$ little cells. A few cells are shaded orange (hit), most are empty (clean). Each cell is an independent coin flip that lands "defect" with tiny probability $p = λ / n$ .

Step 4 — A die survives only if EVERY box is clean

WHAT. The die lives exactly when all $n$ boxes are empty. One box is clean with probability $(1 - p)$ . Because the boxes are independent (specks land independently), we multiply:

$Y = box 1 clean (1 - p) \times (1 - p) \times \dots = (1 - p)^{n} = (1 - \frac{λ}{n})^{n} .$

Term by term:

$1 - p$ = probability one box escapes a defect.
the exponent $n$ = we need all $n$ to escape, so we multiply $n$ copies.
substituting $p = λ / n$ gives the boxed expression in terms of the single number $λ$ .

WHY multiply? "AND" for independent events means multiply their probabilities. "Box1 clean AND box2 clean AND … AND box $n$ clean" → product of $n$ equal factors → a power.

PICTURE. The figure lines up $n$ green "clean" tiles; the survival chance is the product running left to right. If any single tile flips orange, the whole product collapses toward a dead die.

Step 5 — Let the boxes get infinitely fine: $e$ is born

WHAT. Our boxes were an approximation (a defect could in principle share a box). To make the picture exact, shrink the boxes to zero size — that means $n \to \infty$ . We need the limit

$Y = lim_{n \to \infty} (1 - \frac{λ}{n})^{n} .$

WHY a limit? A limit is the tool that answers "what value does this approach as I make the approximation perfect?" Here it removes the artificial grid and gives the true, continuous answer.

The one fact we need. There is a famous limit that defines the number $e \approx 2.71828$ :

$lim_{n \to \infty} (1 + \frac{x}{n})^{n} = e^{x} .$

Here $x = - λ$ . Plugging in:

$Y = e^{- λ} = e^{- D_{0} A}$

Symbol check on the result:

$e$ — the special growth constant this limit hands us.
$- λ = - D_{0} A$ — a negative exponent, because more expected defects means less survival. More dust → smaller $Y$ .

PICTURE. The figure plots $(1 - λ / n)^{n}$ for $n = 1, 2, 4, 10, 50$ climbing toward a dashed horizontal line at $e^{- λ}$ . You literally watch the staircase settle onto the exponential value.

Step 6 — The shape of $Y$ vs area: why big dies are punished

WHAT. Look at $Y = e^{- D_{0} A}$ as a curve in $A$ (with $D_{0}$ fixed). It starts at $Y = 1$ when $A = 0$ and decays, never touching zero.

WHY exponential, not $1/ A$ ? Because doubling area means multiplying the survival factor by itself:

$e^{- D_{0} (2 A)} = (e^{- D_{0} A})^{2} .$

A die twice as big doesn't have half the yield — it has the yield squared. Squaring a number below 1 shrinks it far faster than halving. That's the whole reason giant GPU dies bleed yield.

PICTURE. The plum curve $e^{- D_{0} A}$ next to a dotted " $1/ A$ " curve people wrongly expect. See how the true curve dives faster and the two disagree sharply — the classic mistake exposed.

Step 7 — Edge & degenerate cases (the formula must survive all of them)

We test $Y = e^{- D_{0} A}$ at every extreme so no reader ever hits an unshown scenario.

Case	Input	$Y = e^{- D_{0} A}$ gives	Sanity
Perfect process	$D_{0} = 0$	$e^{0} = 1$	Every die lives ✅
Zero-area die	$A = 0$	$e^{0} = 1$	A point catches no dust ✅
Filthy / huge	$D_{0} A \to \infty$	$e^{- \infty} = 0$	Nothing survives ✅
Small $λ$	$D_{0} A ≪ 1$	$\approx 1 - D_{0} A$	≈ "few defects, small loss" ✅
Defects clump	(Poisson too pessimistic)	use $(1 + \frac{D _{0} A}{α})^{- α}$	reduces to $e^{- D_{0} A}$ as $α \to \infty$ ✅

WHY the last two matter.

For small $λ$ , the exponential's opening slice is $e^{- λ} \approx 1 - λ$ : intuitive "lose roughly $λ$ fraction." Good gut-check for mature nodes.
For clustering, real dust arrives in clumps, so a few dies eat many defects and more dies stay clean than pure randomness predicts — the negative-binomial model (see Poisson distribution) raises $Y$ . As its clustering knob $α \to \infty$ (clumping vanishes) it collapses back to our $e^{- D_{0} A}$ , confirming Poisson is the no-clustering limit.

PICTURE. Three miniature panels: $D_{0} = 0$ (all-green die), $A \to 0$ (a dot, always clean), and clustered dust (orange dots bunched in one corner sparing the rest).

The one-picture summary

The whole derivation on a single canvas: die → tiny boxes → "all clean" product → limit → $e^{- D_{0} A}$ curve, with the edge anchors marked ( $A = 0 \Rightarrow Y = 1$ , big $A \Rightarrow Y \to 0$ ).

Recall Feynman: the walkthrough in plain words

A chip is a rectangle; killer dust specks land on the wafer at random. To find the chance our chip escapes all dust, we chop it into a zillion tiny boxes, each so small it can only hold one speck or none. Each box stays clean with chance $1 - λ / n$ , where $λ = D_{0} A$ is the average number of specks expected on the chip. The chip lives only if every box is clean, so we multiply that chance by itself $n$ times: $(1 - λ / n)^{n}$ . Now make the boxes infinitely tiny ( $n \to \infty$ ) — and this exact expression is the famous definition of $e^{- λ}$ . So $Y = e^{- D_{0} A}$ falls straight out. Because it's an exponential, doubling the chip's area doesn't halve survival — it squares it, which is why huge chips die so easily. Set $D_{0} = 0$ or $A = 0$ and you get $Y = 1$ (nothing to kill it); crank the dust up and $Y \to 0$ . If dust clumps together, a few chips soak up the mess and more survive than pure randomness says — that's the negative-binomial correction, and it quietly turns back into our formula when the clumping disappears.

Recall

Where does the number $e$ enter the yield derivation? ::: From the limit $lim_{n \to \infty} (1 - λ / n)^{n} = e^{- λ}$ when the die is sliced into infinitely many tiny clean-or-not boxes. Why is survival multiplied across boxes? ::: The die lives only if ALL boxes are clean, and independent "AND" events multiply their probabilities. What does doubling die area do to yield? ::: Squares the survival factor: $e^{- D_{0} (2 A)} = (e^{- D_{0} A})^{2}$ — much worse than halving. What is $λ$ and can it exceed 1? ::: The expected count of defects on one die, $λ = D_{0} A$ ; yes, it can exceed 1 — it is a count, not a probability.

Connections

Parent — Yield, defect density, and binning
Poisson distribution — the tiny-boxes limit is the Poisson $P (0)$ ; this page derives it by hand.
Process node scaling — smaller nodes change $A$ and $D_{0}$ , shifting $Y$ .
Chip economics and cost per transistor — $Y$ feeds cost per good die.
Chiplets and MCM — splitting a big die into small ones fights the exponential area penalty shown in Step 6.