Chiplets and multi-die integration: Foundations — Hardware deep dive

This page assumes nothing. Before you read the parent note, you should be able to point at every symbol below and say what it is, what it looks like, and why the topic needs it. Let's earn each one.

1. The wafer, the die, and area $A$

Figure — Chiplets and multi-die integration

Look at the figure. The big circle is the wafer. The grid of squares are the dies. When we say die area $A$ , we mean the size of one of those squares, measured in square millimetres ( $mm^{2}$ ).

The symbol:

$A$ ::: die area, in $mm^{2}$ . Picture: the size of one rectangle in the grid.

2. Defects and defect density $D_{0}$

Defects are scattered across the wafer roughly at random and roughly evenly — like raindrops hitting a pavement. Some areas catch more, some fewer, but on average the sprinkle is uniform.

Why do we need a density and not just a count? Because a count depends on how big a patch you look at. Density lets us ask: "given a die of area $A$ , how many defects do I expect on it?" — and the answer is simply the sprinkle rate times the area.

The symbols:

$D_{0}$ ::: defect density, average defects per $mm^{2}$ . Picture: how thickly the raindrops fall.
$D_{0} \cdot A$ ::: the expected number of defects on a die of area $A$ . Picture: raindrops on one square = rate × square-size.

3. Probability, and why "zero defects" is the goal

A die works only if it catches zero defects. So we need to talk about chance.

Here the "something" we care about is how many defects land on one die. We'll write $P (k)$ = the probability that exactly $k$ defects land on a die, where $k = 0, 1, 2, \dots$

$k$ ::: a whole-number count of defects on one die ( $0, 1, 2, \dots$ ).
$P (k)$ ::: the chance of getting exactly that many.

Why zero specifically? Because one defect is (usually) enough to kill the die. So the die is good exactly when $k = 0$ . The chance of that is the number we're after — and we're about to name it yield.

4. Building the yield formula, piece by piece

We want one number: the probability a die has zero defects. We'll call that number the yield $Y$ .

$Y$ ::: yield — the fraction of dies that come out good; equivalently the probability that one die has zero defects (a number from $0$ to $1$ ).

To get $Y$ we first need the full probability $P (k)$ for any count $k$ , then set $k = 0$ . That full expression — the Poisson formula — uses two pieces of notation a newcomer won't recognise: the factorial $k!$ and the exponential $e^{x}$ . Let's earn each, then assemble them and read off $Y$ .

4a. Factorial $k!$

Why does it appear? When you sprinkle $k$ identical defects onto a die, the order you dropped them in doesn't matter — "defect A then B" is the same outcome as "B then A". There are $k!$ ways to order $k$ things, so if we counted ordered sprinklings we'd have over-counted each real outcome exactly $k!$ times. Dividing by $k!$ cancels that over-count. (For our purpose you only need one fact: $0! = 1$ , the value we plug in for a good die.)

4b. The exponential $e^{x}$

Why this function and not, say, $1/ x$ or a straight line? Because we're asking: "if defects land independently, what's the chance none of them hits my die?" Independent survival multiplies. Split the die into many tiny patches; the chance the whole die survives is the product of each patch surviving. When you multiply many independent tiny survival-chances together, the maths forces an exponential to appear — that's exactly what $e^{- D_{0} A}$ is. No other function fits "independent, multiplying survival."

Look at the red curve: as $D_{0} A$ (the expected defects) grows along the horizontal axis, survival $e^{- D_{0} A}$ falls off a cliff. That cliff is the reason big dies are doomed.

4c. Assembling the Poisson formula — where each factor comes from

The full probability of exactly $k$ defects on a die is:

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

Read it as three factors, each with a job:

$(D_{0} A)^{k}$ — "each defect had a chance to land here." We established $D_{0} A$ is the expected count of defects on this die. Landing $k$ of them means multiplying that single-defect tendency by itself $k$ times → the power $(D_{0} A)^{k}$ . More expected defects, or more of them wanted ( $k$ ), pushes this factor up.
$\frac{1}{k !}$ — "don't double-count orderings." From §4a: the $k$ defects are identical, so we divide out the $k!$ equivalent orderings we'd otherwise over-count.
$e^{- D_{0} A}$ — "and all the rest of the area stayed empty." This is the survival factor from §4b: the probability that no extra defects landed. It's the exponential we argued must appear whenever independent survival-chances multiply.

Multiply the three together and you have the standard Poisson model — the textbook description of rare, independent, uniformly-sprinkled events.

Now the payoff. Set $k = 0$ for a good die. Then $(D_{0} A)^{0} = 1$ (factor 1 vanishes — no defects to place), $0! = 1$ (factor 2 is trivial), leaving only the survival factor:

$Y = P (0) = e^{- D_{0} A}$

5. Splitting a die: the symbol $N$ (and the catch)

If the big die was $800 mm^{2}$ and $N = 4$ , each chiplet is $800/4 = 200 mm^{2}$ . Plug the smaller area into the yield formula:

$Y_{chiplet} = e^{- D_{0} (A / N)}$

A smaller exponent means the exponential is closer to 1 — each individual chiplet is far likelier to be good. So far so good.

Common mistake "Splitting into N chiplets automatically raises the yield of the finished product."

Why it feels right: each small piece has higher yield, so the whole thing must too. The catch: a package only works if all $N$ chiplets inside it work. If they're independent, the chance all survive is each chiplet's yield multiplied $N$ times: $Y_{package} = (e^{- D_{0} A / N})^{N} = e^{- D_{0} A}$ — exactly the monolithic yield again. Cutting alone moves the risk, it doesn't erase it. Fix — where the real win comes from: ① Known-good die (KGD): test every chiplet before assembly and discard the bad ones, so you never pay to package doomed silicon (you save cost, and only assemble survivors). ② Redundancy / spares / binning: build a few extra chiplets or extra cores and disable the faulty ones, so a defect degrades rather than kills the product. The exponential is why splitting is necessary; KGD and redundancy are why it actually pays.

So the honest summary: splitting shrinks each piece's risk and — crucially — lets you filter before you glue. That filtering, not the split by itself, is the economic engine.

6. The energy symbols: $E_{bit}$ , $C$ , $V$ , $L$

The parent's second formula talks about the cost of moving data between dies. Four new symbols:

The picture: think of a wire as a bucket. To signal a "1" you fill the bucket with charge; to signal "0" you empty it. A longer wire is a bigger bucket — $C \propto L$ (capacitance grows with length). Every bit you send costs energy to fill that bucket:

$E_{bit} = \frac{1}{2} C V^{2}$

Why does this matter for chiplets? Dies glued close together (micrometres apart, on the package) have tiny buckets → cheap bits. Data forced off the package onto a circuit board (centimetres) has huge buckets → expensive bits. Closeness = cheap communication. That's why we bring dies physically near each other.

The symbol $\propto$ ::: "is proportional to" — grows in step with, ignoring constants. $C \propto L$ means "double the length, double the capacitance."

Where to go next

Wafer yield and defect density — the deep dive on $Y = e^{- D_{0} A}$ , defect clustering, and how $D_{0}$ is measured.
Moore's Law and its slowdown — why we can't just keep shrinking one big chip.
Advanced packaging (2.5D 3D interposers) — how the good dies get glued together.
Memory bandwidth and HBM — where the $E_{bit} \propto L$ argument pays off hugely.
Interconnect energy and dark silicon — the energy side, expanded.
Heterogeneous integration — mixing different dies in one package.
Back to the parent: Chiplets and multi-die integration.

Equipment checklist

Test yourself — you're ready when each answer comes instantly.

What does the symbol

A

stand for, and its units?

Die area, in

mm^{2}

— the size of one chip rectangle.

What does

D_{0}

measure?

Defect density: average number of defects per

mm^{2}

of silicon.

What is the pure number

D_{0} A

The expected (average) number of defects landing on one die of area

A

What does

P (k)

mean?

The probability that exactly

k

defects land on a die.

In the Poisson formula, what jobs do

(D_{0} A)^{k}

and the

1/ k!

do?

(D_{0} A)^{k}

= the tendency for

k

defects to land;

1/ k!

cancels the over-counted orderings of identical defects.

Compute

0!

and

e^{0}

0! = 1

and

e^{0} = 1

e^{- 4}

big or small, roughly?

Small — about

0.018

;

e^{- x}

shrinks toward zero as

x

grows.

Why must yield be an exponential

e^{- D_{0} A}

and not a straight line?

Independent survival of many patches multiplies, and multiplying independent survival chances forces an exponential.

Which value of

k

gives a good die, and what does the formula reduce to?

k = 0

; it reduces to

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

Name two assumptions the Poisson yield model makes, and one way reality breaks them.

Uniform density and independent defects; reality breaks both via clustering (defects spawn neighbours) and wafer non-uniformity.

If you split area

A

into

N

equal chiplets, what is the yield that all survive, and why doesn't splitting alone help?

(e^{- D_{0} A / N})^{N} = e^{- D_{0} A}

— the same as monolithic; the real gain comes from testing/discarding bad dies (KGD) and redundancy.

State the energy-per-bit formula and what

C \propto L

means.

E_{bit} = \frac{1}{2} C V^{2}

; capacitance grows with wire length, so longer wires cost more energy per bit.

What does the symbol

\propto

mean?

"Is proportional to" — grows in lockstep, ignoring constant factors.