yield

Yield is the fraction of manufactured devices that work, usually expressed as the percentage of dies on a wafer that pass testing, and it is the number that decides whether a semiconductor product makes money. A wafer costs about the same to process whether its chips are good or bad, so every failed die is paid for by the good ones. In the earliest transistor years the usable fraction of a production run was often small and prices were correspondingly high, and the history of the industry can be read as a long campaign to push yield upward, until billions of transistors could be sold for a few currency units per chip.

Die yield is set mostly by defects and die area, and the relationship is brutal. A defect, such as a particle, a scratch, a pinhole in an oxide or a crystal flaw, kills a chip if it lands in a critical place. The simplest description, the Poisson model, treats defects as randomly scattered across a wafer with an average density D per square centimetre, so a die of area A has a yield of e raised to the power of minus A times D; at one defect per square centimetre, a one-square-centimetre die yields about 37 percent, while a die a quarter of that size yields about 78 percent. The Murphy model, from a 1964 analysis of the cost-size trade-off for monolithic circuits, allows defect density to vary across a wafer and predicts somewhat higher yields for large dies, and later refinements fit real fabs better still. In words, big chips and dirty processes compound each other's losses.

Early transistor makers lived with low yields, and that explains the prices. Germanium and then silicon transistors of the 1950s were grown or alloyed one at a time, with hand-controlled steps and contamination that was poorly understood, so many units failed to meet their specification or fell into lower-grade bins. A part of the output was often sold as a cheaper grade rather than discarded, which is why early catalogues had the same device number at several prices. Texas Instruments' announcement of the first commercial silicon transistor on 14 May 1954 was as much a manufacturing achievement as a scientific one, since silicon was harder to purify and process reliably than germanium. Costs fell as crystal growth, diffusion and cleanliness improved, and the planar process of 1959 helped by leaving the junctions covered with oxide, reducing surface contamination and lifting yields.

Yield drives integrated-circuit economics because it ties cost to chip size. The cost of one good chip is the cost of a processed wafer divided by the number of good dies on it, so doubling die area both halves the number of candidates and lowers the probability that each is good, which is why price rises faster than size. This is the economic reason designers make memory and logic ever more compact, and why a modest scaling gain can pay for itself. It also explains the practice of adding redundancy: memories carry spare rows and columns that can be switched in to replace defective ones, and large processors are sold in several grades when some blocks fail, turning near-misses into saleable products.

Wafer testing and yield learning turn failures into process improvements. After fabrication, probes contact each die in a step called wafer sort, marking good and bad units on a map. Patterns in that map point to causes: failures concentrated at the edge suggest edge effects, clusters suggest contamination or a scratch, repeating patterns suggest the mask or a tool. Engineers feed this back to the line in what the industry calls yield learning, and a new process normally starts with low yield and climbs over months as problems are found. A distinction worth keeping is that die yield is not the only yield: line yield counts wafers lost in processing, and final test yield counts packaged parts that fail after assembly, so the total yield is the product of these.

A common misconception is that yield is simply a matter of dust and clean rooms. Particles certainly matter, which is why fabs use filtered air and strict procedures, but a mature process can also lose dies to subtle parametric problems: a transistor that works but is too slow or too leaky is a failure even though nothing is broken. Another pitfall is quoting a single percentage without saying what it measures, since a headline yield may exclude packaging or test loss. Modern chiplet designs partly exist to exploit this arithmetic, because several small dies each yield far better than a single huge one, even after the cost of joining them.

Read yield as the quiet number behind every chip price: a probability that each die survives the defects scattered across a wafer, so that smaller area, cleaner processing and learning from failures matter as much to cost as clever circuit design.