Chiplets and multi-die integration: Visual walkthrough

Prerequisites we lean on: Wafer yield and defect density and the parent topic. This is the picture-first companion to Wafer yield and defect density.

Step 1 — The wafer and the raisins

WHAT: We picture defects as random dots scattered over the wafer.

WHY: Before any math, we need a mental model. If a single defect lands inside a chip, that chip is (in the simple model) dead. So the whole story reduces to one question: how likely is a chip to have zero dots on it?

PICTURE: Below, the peach disc is the wafer; the magenta dots are defects sprinkled at random.

Figure — Chiplets and multi-die integration

Step 2 — Counting defects on one chip: the expected number

WHAT: We form the single most important quantity, the expected defect count on one die:

$expected # defects λ = defects per mm^{2} D_{0} \times die area in mm^{2} A$

Term by term: $D_{0}$ is how thick the sprinkle is, $A$ is how much area you exposed to the sprinkle, and their product $λ$ (Greek letter "lambda") is how many defects you should expect to land, on average.

WHY $λ$ and not the actual count? Because the actual count is random — sometimes $0$ , sometimes $3$ . We can't predict a single wafer, but we can predict the average. $λ$ is the anchor everything else hangs on.

PICTURE: Two chips of different area over the same sprinkle. The bigger one is expected to catch more dots. That is the whole intuition for "big dies are doomed."

Step 3 — Why the Poisson formula? (the tool, justified)

WHAT: We write the probability of getting exactly $k$ defects on a die whose expected count is $λ = D_{0} A$ :

$P (k) = \frac{( D _{0} A ) ^{k} count ^{k} e ^{- D_{0} A} decay}{k factorial k !}$

Reading it left to right:

$(D_{0} A)^{k}$ — the expected count raised to the power $k$ : bigger $λ$ makes larger $k$ more likely.
$e^{- D_{0} A}$ — a shrinking factor ( $e \approx 2.718$ ; a negative exponent means "less than one"). It keeps the total probability adding to exactly $1$ .
$k!$ — " $k$ factorial" $= k \times (k - 1) \times \dots \times 1$ . It corrects for the many orderings the same $k$ defects could arrive in.

WHY this shape? Don't memorise it — just trust the three assumptions force this exact form. We only need one value of $k$ next.

PICTURE: The Poisson bar chart for $λ = 1$ and $λ = 4$ . Notice: as $λ$ grows, the bar at $k = 0$ (our "good chip" bar, highlighted) shrinks fast.

Step 4 — A good chip means ZERO defects: plug in $k = 0$

WHAT: Substitute $k = 0$ into the Poisson formula:

$Y = P (0) = \frac{( D _{0} A ) ^{0} e ^{- D_{0} A}}{0 !}$

WHY these two simplifications:

$(D_{0} A)^{0} = 1$ — anything raised to the power zero is $1$ .
$0! = 1$ — factorial of zero is defined as $1$ .

Both the top-left and the bottom collapse to $1$ , leaving only the decay factor:

$Y = e^{- D_{0} A}$

PICTURE: The full curve $Y = e^{- D_{0} A}$ against area $A$ , with the $k = 0$ bar from Step 3 traced onto it — showing that the whole yield curve is literally the height of the "zero defects" bar as $A$ slides.

Step 5 — Feel the exponential: the monolithic die

WHAT: Take the parent's numbers: $D_{0} = 0.005$ defects/mm² and a monolithic die of $A = 800$ mm².

$Y = e^{- 0.005 \times 800} = e^{- 4} \approx 0.018$

Term by term: $0.005 \times 800 = 4$ expected defects per die. Four expected defects means the chance of zero is $e^{- 4} \approx 1.8%$ . Ninety-eight of every hundred giant dies are born dead.

WHY show this first? So the rescue in Step 6 has something brutal to rescue us from.

PICTURE: One huge die outline over the sprinkle — almost every copy has at least one dot inside. The rare survivor is circled.

Step 6 — The rescue: split into $N$ chiplets

WHAT: The yield of one chiplet of area $A / N$ :

$Y_{chiplet} = e^{- D_{0} \cdot A / N}$

The only change from Step 4 is the exponent: $D_{0} A$ became $D_{0} A / N$ — a smaller expected count, so a bigger survival probability.

With $N = 4$ , $A / N = 200$ mm²:

$Y_{chiplet} = e^{- 0.005 \times 200} = e^{- 1} \approx 0.368$

Now $0.005 \times 200 = 1$ expected defect per chiplet, so $\approx 36.8%$ survive — about 20× better than the $1.8%$ monolith.

WHY it's more than fairness: We never build the 98%-doomed giant. We keep only the good small dies (known-good die), then glue those into one package.

PICTURE: The same sprinkle, now with the die cut into 4 tiles. Bad tiles (with a dot) shaded out, good tiles glowing. Most tiles survive.

Step 7 — Edge case: when splitting does NOT help

WHAT: Compare a small die, say $A = 50$ mm², $D_{0} = 0.005$ :

$Y = e^{- 0.005 \times 50} = e^{- 0.25} \approx 0.779$

Already $78%$ good. Split it in two ( $A / N = 25$ ): $Y_{chiplet} = e^{- 0.125} \approx 0.882$ . A modest gain — and now you owe interposer + assembly + double the I/O area. Chiplets can end up more expensive here.

WHY include this? The contract: cover every case. Chiplets are not free magic. The Wafer yield and defect density curve is flat near $A = 0$ , so the payoff lives only where the curve is steep — large $A$ .

PICTURE: The $Y$ – $A$ curve annotated with two regions: a flat "splitting barely helps" zone near the origin (violet) and a steep "splitting rescues you" zone at large area (magenta).

Step 8 — Degenerate cases: perfect and worst wafers

WHAT: Push the model to its edges:

Input	Plug in	Result	Meaning
$D_{0} = 0$ (flawless wafer)	$e^{- 0 \cdot A} = e^{0}$	$Y = 1$	Every die good — splitting pointless.
$A \to 0$ (vanishing die)	$e^{- D_{0} \cdot 0} = e^{0}$	$Y \to 1$	Infinitely small dies never catch a defect.
$A \to \infty$ (giant die)	$e^{- D_{0} \cdot \infty}$	$Y \to 0$	Nothing survives — the reticle limit and yield both scream "stop."

WHY: These aren't trivia — they're the sanity checks that prove $e^{- D_{0} A}$ behaves like real silicon. A formula that gave $Y = 5$ or $Y = - 0.3$ would be nonsense; ours always sits in $[0, 1]$ .

PICTURE: The single curve $Y = e^{- D_{0} A}$ with its two flat asymptotes: hugging $Y = 1$ on the left, hugging $Y = 0$ on the right, never leaving the $[0, 1]$ band.

The one-picture summary

Recall Feynman retelling of the whole walkthrough

Sprinkle raisins on cookie dough — you can't aim them. One raisin ruins the cookie it lands in. If you bake one enormous cookie ( $A = 800$ ), it's almost certain to catch a raisin, so almost every giant cookie is trash — we did the sum: only $1.8%$ survive ( $e^{- 4}$ ). The math that says so is just "how likely is zero raisins?", and for random sprinkles that answer is $e^{- D_{0} A}$ — an exponential, meaning bigger cookies get punished faster and faster. Now bake four small cookies ( $A /4 = 200$ ) from the same dough: each catches raisins only a quarter as often, so $36.8%$ survive ( $e^{- 1}$ ), and we taste-test each one and toss the duds before gluing the good ones together. But if the cookie was already tiny ( $A = 50$ ), it was $78%$ fine anyway, so splitting barely helps and you just pay extra for the glue. And the extremes behave: no raisins at all → everything survives; a cookie of zero size → always fine; an infinite cookie → always doomed. That last curve — starting at $1$ , sliding down to $0$ , never going below zero or above one — is the whole story of why we build chiplets.

Connections

Wafer yield and defect density — the home of the $e^{- D_{0} A}$ curve we rebuilt here.
Advanced packaging (2.5D 3D interposers) — the cost you pay once you split into chiplets.
Moore's Law and its slowdown — why big monolithic dies stopped scaling.
Heterogeneous integration — mixing process nodes, another reason to split.