Yield Learning Rate
Yield learning rate describes how die yield improves as cumulative production volume grows, typically following a learning-curve pattern where defect density falls by a roughly fixed percentage — often somewhere around 20 to 40 percent — for every doubling of cumulative wafers produced. It's a distinctly different way of looking at the same underlying ramp as yield ramp slope: rather than tracking yield against calendar time, learning rate tracks yield against accumulated production experience itself, which turns out to be the more fundamental driver.
The underlying power-law learning model treats defect density as a function of cumulative wafer count rather than of elapsed calendar time, which is exactly what distinguishes learning rate as a concept from yield ramp slope — the same physical fab running the same process at double the wafer-start rate would show the same learning-rate improvement in half the calendar time. This cumulative-volume framing captures something genuinely different from a simple time-based ramp curve: it reflects the idea that yield improvement comes from accumulated process knowledge and experience, which depends on how much the process has actually been run, not on how many days have simply elapsed since it started.
The doubling rule — each doubling of cumulative production reducing defect density by a roughly fixed fraction, commonly cited in the 20 to 40 percent range — gives learning rate a characteristic diminishing-returns shape, since doublings of cumulative volume take progressively longer to achieve as volume grows. Going from 1,000 to 2,000 cumulative wafers is one doubling; going from 100,000 to 200,000 is also one doubling but represents vastly more additional wafers and additional calendar time to achieve. This is exactly why learning curves characteristically show rapid early improvement that then visibly slows, even though the underlying doubling-based improvement rate itself hasn't actually changed.
Learning is fastest early in a technology's production life because the first few doublings of cumulative volume are precisely when the biggest, most obvious yield-loss mechanisms get identified and fixed — later doublings mostly chase progressively smaller, harder-to-find remaining mechanisms, producing visibly diminishing absolute gains even at a constant underlying learning-rate percentage. A process early in its life typically still has large, systematic, easily-diagnosed yield-loss mechanisms present, and finding even one of these produces a large defect-density improvement. As those get progressively eliminated, what remains is smaller, more subtle, harder-to-isolate mechanisms — meaning each subsequent doubling of volume produces a smaller absolute yield gain even while following the same percentage-per-doubling learning rate.
Every learning curve eventually flattens into a mature yield ceiling, set by the process's fundamental capability rather than by any remaining fixable defect mechanism — once that ceiling is reached, additional cumulative volume stops producing further defect-density improvement no matter how many more doublings accumulate. This ceiling represents the point where learning-driven improvement genuinely runs out, distinct from the diminishing-but-still-positive gains seen earlier in the curve — further yield improvement beyond this ceiling requires a genuine change to the underlying process capability itself, not simply more accumulated production experience.
Learning rate projections feed directly into fab capacity investment decisions and competitive positioning, because the speed at which a given foundry's process learns determines how quickly that capacity investment starts generating a return, and a demonstrably faster learning rate than competitors is a genuine, quantifiable competitive advantage. A foundry able to show that its process learns faster — reaching a given target yield after fewer cumulative wafers than a competitor's comparable process — can make a stronger case to customers about predictable near-term yield and cost trajectory, which is exactly why learning rate shows up as a tracked metric in technology-node business planning rather than as a purely academic curiosity.