Defect Density Console

Strip a fab's economics down to one number and it's the defect density. Yield — the fraction of dies that work — is what turns wafers into sellable chips, and yield is governed, above all, by how many killer defects land per unit area. The relationship is exponential: yield equals e to the minus area times D0, so every reduction in defect density lifts yield disproportionately, and every increase in die area makes the part more sensitive to the same defects.

The simplest model treats defects as randomly scattered (Poisson), but reality is kinder than that. Defects cluster — a particle shower or a process non-uniformity concentrates damage in one region, killing dies that would otherwise have been spread across the wafer, so fewer total dies are lost. The negative-binomial model captures this with a clustering factor, and it consistently predicts higher, more realistic yields than pure Poisson. Seeing both, as this console shows, keeps you from being needlessly pessimistic.

Defect density isn't static — it falls. A new process node launches with high D0 and poor yield, and then the fab climbs the yield-learning curve: engineers find defect sources, eliminate them, and D0 drops year over year, often steeply at first. This trajectory is the difference between a product that loses money at launch and the same product comfortably profitable once the node matures, which is why the learning-curve projection here is as important as the instantaneous yield.

And because yield depends on area times D0, large dies are punished hardest — the same defect density that barely dents a small controller can halve a reticle-sized AI die. That sensitivity is why die size is a first-order cost decision and why chiplets exist. Take the good-dies-per-wafer figure from here into the Wafer Cost console for cost per chip, and compare yield models in the Yield console.