Visual walkthrough — Wafer-scale engines (Cerebras-style)

We build the central result $Y=e^{-DA}$ from nothing, show why it kills a big chip, then derive the redundancy fix $M\ge \frac{N}{1-f}$ that saves it.

---

Step 1 — What a wafer and a defect even are

WHAT: picture the wafer as a flat baking sheet, and defects as raisins sprinkled on it at random.

WHY start here: every later symbol ($D$, $A$, $\lambda$) is just a way of counting these raisins in a region. If you can see the raisins, the algebra is obvious.

PICTURE: the sheet below has defects scattered over it. A small die (dashed box) sits on the sheet — notice it happens to be empty, so it survives. A big die (solid box) covers most of the sheet and swallows several defects, so it dies.

---

Step 2 — Defect density $D$: turning raisins into a number

WHAT: we collapsed "a messy scatter of dots" into one steepness-like number.

WHY this tool and not just "count defects"? A raw count is useless — it depends on how big a region you look at. Density is the count per unit area, so it stays the same whether you look at a tiny die or the whole plate. That is exactly the property we need, because next we will ask "how many defects land in a die of area $A$?" — and density × area answers that instantly.

PICTURE: the same wafer, now with a $1{\text{mm}}^2$ reference square drawn on it. Counting dots inside that unit square is $D$.

---

Step 3 — Expected defects in a die: the product $\lambda =D\cdot A$

$\underset{\text{avg defects in the die}}{\underbrace{\lambda }}=\underset{{\text{defects per mm}}^2}{\underbrace{D}}\times \underset{{\text{mm}}^2\text{ of the die}}{\underbrace{A}}$

WHAT: density times area gives a pure count — the ${\text{mm}}^2$ units cancel, leaving "defects."

WHY: if raisins are spread at $D$ per unit area, then over $A$ units of area you expect $D\cdot A$ of them. This is the same logic as "$3$ cars per km over $10$ km $\Rightarrow$ expect $30$ cars." We call this average $\lambda$ (the Greek letter lambda) because that is the standard name for the mean of the distribution we meet next.

PICTURE: two dies side by side — a small one ($A$ small $\Rightarrow \lambda$ small) and a wafer-sized one ($A$ huge $\Rightarrow \lambda$ huge). The bar under each shows its expected defect count growing with area.

---

Step 4 — Why the Poisson distribution, and why $P(0)$

The right tool is the Poisson distribution: it gives the probability of getting exactly $k$ rare, independent events when you expect $\lambda$ of them.

Now set $k=0$ (we want zero defects):

$Y=P(0)=\frac{{\lambda }^0{e}^{-\lambda }}{0!}=\frac{1\cdot e^{-\lambda }}{1}=e^{-\lambda }=\boxed{e^{-DA}}$

• ${\lambda }^0=1$ (anything to the power $0$ is $1$).
• $0!=1$ (there is exactly one way to arrange nothing).
• So both the top power and the bottom factorial vanish to $1$, leaving the clean $e^{-DA}$.

WHAT: we plugged $k=0$ into the general rule and everything except $e^{-\lambda }$ collapsed.

WHY: "$k=0$" is the engineering question "did this die dodge every defect?" — the yield $Y$.

PICTURE: the Poisson bar chart for a fixed $\lambda$. The single bar at $k=0$ is highlighted — that lone bar is the yield.

---

Step 5 — The yield curve, and the death of the big die

Let's read the curve in three regimes — the all-cases requirement:

• Tiny die ($DA\to 0$): $e^{-\text{(small)}}\approx 1$. Almost every die is perfect. This is why we normally dice a wafer into small chips.
• Medium die ($DA\approx 1$): $Y\approx e^{-1}\approx 0.37$. About a third survive — normal chip economics.
• Whole-wafer die ($DA\gg 1$, e.g. $DA=46$ for a plate-sized die at realistic $D$): $Y=e^{-46}\approx 10^{-20}$. You would need more wafers than atoms in a room to get one perfect one.

WHAT: the exponential turns a large area into a vanishingly small survival chance.

WHY it matters: this is the parent note's "yield killer." A monolithic, defect-free full wafer is impossible. If that were the only option, wafer-scale engines could not exist.

PICTURE: the smooth decay curve $Y=e^{-DA}$ with the three regimes marked by dots. Watch the curve fall off a cliff as we move right into wafer territory.

---

Step 6 — Change the question: tolerate dead cores

Define the new quantities:

• $M$ = number of cores physically built on the wafer.
• $f$ = the fraction of cores expected to be dead (from defects). So $f$ is between $0$ and $1$.
• $N$ = number of good cores we actually need for the product.

The number of surviving cores is the built cores minus the dead ones:

$\text{good cores}=\underset{\text{built}}{\underbrace{M}}-\underset{\text{dead}}{\underbrace{fM}}=M(1-f)$

We just need enough survivors:

$\underset{\text{good cores}}{\underbrace{M(1-f)}}\ge \underset{\text{cores needed}}{\underbrace{N}}$

Solve for $M$ by dividing both sides by $(1-f)$ (which is positive, so the inequality direction is safe):

$\boxed{M\ge \frac{N}{1-f}}$

• $\frac{1}{1-f}$ = the over-provisioning factor: build this many times more cores than you strictly need.

WHAT: we replaced "wafer must be perfect" with "wafer must have $\ge N$ good cores," and got a build recipe.

WHY this tool (a simple ratio, not the exponential): yield $e^{-DA}$ answered "is the whole die perfect?" — the wrong question. The right question is a counting question about survivors, and counting only needs "kept fraction × total," i.e. simple proportions.

PICTURE: a mesh grid where a handful of cores are crossed out (dead). A path threads around them from one side to the other — the reroute. A counter shows good cores still $\ge N$.

---

Step 7 — Reading the over-provisioning factor in all cases

$M\ge \frac{N}{1-f}$

Let's cover every regime of $f$ so no scenario surprises us:

• $f=0$ (no defects): $M\ge \frac{N}{1}=N$. Build exactly what you need — the ideal, never real.
• $f=0.10$ (10% die): $M\ge \frac{N}{0.9}\approx 1.11N$. Build about 11% extra cores as spares.
• $f=0.5$ (half die): $M\ge 2N$. You must build double.
• $f\to 1$ (almost all die): $\frac{1}{1-f}\to \infty$. The factor blows up — wafer-scale becomes hopeless. This is the degenerate limit: redundancy only works when $f$ is small.

WHAT: the ratio told us exactly how many spares to bake in.

WHY it matters: this is why the parent's core count is what it is — the number isn't round marketing; it's $N/(1-f)$ rounded up.

PICTURE: the over-provisioning factor $\frac{1}{1-f}$ plotted against $f$. Flat and cheap near $f=0$, shooting to infinity as $f\to 1$ — the marked points $f=0,0.1,0.5$ show the "cheap zone" WSEs live in.

---

The one-picture summary

This final figure compresses the whole journey: the exponential yield wall on the left (perfection is impossible), the mesh-with-reroute in the middle (change the question), and the gentle over-provisioning curve on the right (a small spare fraction fixes everything).

Recall Feynman: retell the whole walkthrough in plain words

Sprinkle raisins (defects) randomly on a baking sheet (wafer). Count how crowded they are per unit area — that's $D$. Multiply by how big your cookie is ($A$) and you get how many raisins you expect inside it, called $\lambda$. But you only want cookies with zero raisins, and the chance of that — for independent sprinkles — is $e^{-\lambda }=e^{-DA}$. For a small cookie that's nearly $1$ (easy). For a cookie the size of the whole sheet, $\lambda$ is huge and the chance is basically zero — a perfect giant cookie never happens. The trick: don't bake one giant cookie, bake one giant sheet of tiny connected cells. A few cells get raisins and die, but the roads between cells let you skip the dead ones. Now you just need "enough living cells," so you bake a few extra: build $M=N/(1-f)$ cells, and if only a small fraction $f$ die you still have your $N$ good ones. Physics said "impossible"; changing the question to "enough good cores" made it easy.

Recall Check yourself

Why does the Poisson factor $e^{-\lambda }$ appear instead of some other function? ::: Because defects land independently and memorylessly; independent random arrivals always accumulate the factor $e^{-\lambda }$, the fingerprint of "each tiny region independently might get hit." Why is $Y=e^{-DA}$ the wrong metric for a wafer-scale engine? ::: It measures a defect-free whole die; WSEs never need perfection, only enough good cores, so the right metric is the survivor count. If $f=0.2$, how many cores must you build to get $N$ good ones? ::: $M\ge N/(1-0.2)=1.25N$, i.e. 25% extra.

Related: Yield & Defect Density Models · Dennard Scaling & the Memory Wall · Network-on-Chip (NoC) & Mesh Topologies · Chiplets & 2.5D/3D Integration · Systolic Arrays & TPUs · Sparsity in Neural Networks · Liquid Cooling & Power Delivery Networks