The Economics of Chiplets — We Split Dies Because We Cannot Build Them Big

Chips are split into chiplets not because splitting is faster but because a single large die cannot be built at a price anyone will pay. The reticle limit, the exponential in yield, cost per good die, the point where splitting starts to lose money, mixing process nodes, and UCIe as a standard for the seam.

Splitting Does Not Make It Faster

Cutting a design into several small dies and putting them in one package — a chiplet build looks like a performance technique. It is the opposite. Wires are always faster and cheaper in power when you do not split them. If the whole design fits on one die, that version wins on every technical axis. We split anyway, because the single-die version cannot be built at a price anyone will pay.

So most of the chiplet argument is arithmetic, not physics. What does one working die cost? How many products come off one wafer? How many products can share the fixed cost of a design? How the pieces get sewn back together — interposers, HBM, hybrid bonding — is covered in Advanced Packaging. This article looks only at why splitting pays, and where it stops paying.

An Analogy: One Big Pane of Glass, or Cut Panes

Picture a factory making large panes of glass. Now and then a small bubble gets into the material. Where it lands is random, and you cannot fully prevent it.

• A pane with even one bubble is scrap
• The larger the pane, the lower its chance of containing no bubbles at all
• Cut the sheet down before inspection and you scrap only the one small pane that caught the bubble

That is the whole economic core of chiplets. The bubbles are defects landing on a wafer; the size of the pane is die area. Building one giant die is the business of throwing everything away over a single bubble, and splitting simply makes the unit you throw away smaller.

One thing differs from glass. Cut panes sell as they are; chips do not work until you sew them back together. The seam — the circuit area that links dies, the power to cross a boundary, the assembly work — grows with every extra piece. So this technique was always a scale with yield savings on one side and seam costs on the other. Which way it tips depends on area and defect density.

The Ceiling Does Not Move

Why can't we build big in the first place? A scanner can only print so much area in one exposure — roughly 26 mm × 33 mm on current tools, about 858 mm². A circuit above that reticle limit cannot exist as a single printed pattern.

And the ceiling is moving down, not up. Next-generation High-NA EUV halves the printable field for optical reasons, to roughly 26 mm × 16.5 mm. The further we push scaling, the less circuit fits in one pattern — scaling and splitting advance together rather than compete (EUV Lithography).

In practice, though, yield bites before the reticle limit does. Even at buildable sizes, the fraction of good dies falls exponentially with area. Get a feel for that exponential first.

What Does One Good Die Cost?

The unit that matters is not the price of a wafer, nor the area of a die, but the price of one working die. Two numbers set it: how many dies you can cut from a wafer ($N$), and what fraction of them work ($Y$).

$C_w$ is the price of one wafer and $A$ is the area of one die. Read it as "wafer cost divided by the number of chips you can actually sell." Put in words: if half the dies on a wafer come out dead, each surviving die carries twice the wafer price. The cost of the scrapped half does not disappear — it is billed to the half that shipped. The whole split-or-not decision is about how that denominator moves.

Of the two, $Y$ becomes an exponential once you assume defects land at random.

$D_0$ is the defect density per unit area. Read it as "increase the area and the good-die fraction falls exponentially" — which says, in words, that doubling the area does not halve the surviving fraction but squares it. A part yielding 80% drops to about 64% at twice the area and about 41% at twice that again. Every increment of area multiplies the survivors down rather than subtracting from them. That single line is the entire motivation for splitting.