DIES PER WAFER

Big squares on a round wafer

⚪

A 300 mm wafer has about 70,000 mm² of area. Big square dies waste more of the round edge, so a wafer yields only about 60 to 70 candidate dies of H100 size, compared with hundreds of smaller phone chips.

Before yield even enters the picture, big dies get fewer chances per wafer.

YIELD

Defects land at random

Tiny particles and pattern faults land randomly across the wafer. A die is good only if no killer defect lands on it. In a simple Poisson model, yield = e^{−D × A}, where D is the defect density and A is the die area.

Double the area and yield doesn't halve; it is squared, falling much faster.

FIGHTING BACK

Spares and harvesting

🧩

Redundancy

Build in spare SMs or cores; the H100 die has 144 SMs but ships with at most 132 enabled

🏷️

Binning

Sell dies with a few faults as lower-tier products

🧪

Known-good die

Test every chiplet before packaging so one bad die doesn't waste a package

Almost no large AI die is perfect; it is designed to work anyway.

THE BILL

What goes into the price

💵

Wafer Press reports put a 3 nm wafer at around $18,000–$20,000 and a 2 nm wafer at up to $30,000; foundries don't publish prices

Dies Only the good dies from each wafer can be sold

HBM Often the single biggest part of the bill of materials

Packaging and test CoWoS, assembly and burn-in add a large share

Manufacturing cost is only part of the price: years of design and software are paid for too.

DESIGN COST

Paying for the design

📐

Designing a leading-edge chip can cost hundreds of millions of dollars in engineers, software tools, masks and test runs. A full mask set alone for the newest nodes costs many millions. That cost is spread over every chip sold, so only high-volume or high-value chips justify it.

This is why only a handful of companies can afford to build leading AI chips.

UNIT 13 STUDY COMPLETE

🎯

Ready for the Fab Challenge?

You've seen how defect density and die size set the yield, and why big AI dies are so expensive.