Sampled Wafer Test
Sampled wafer test is a production testing strategy that tests only a statistically chosen subset of die on a wafer rather than every single die, trading some risk of an untested defective die shipping against meaningful reductions in test time and cost — a tradeoff that only makes economic sense once a product's yield and risk profile genuinely justify accepting it. The decision to sample rather than fully test isn't a cost-cutting shortcut applied indiscriminately; it's a deliberate statistical risk-management decision that has to be justified by specific, quantifiable conditions around yield level, safety criticality, and margin to specification.
Grid sampling tests die at fixed, evenly spaced locations across the wafer, which assumes that defects of concern distribute roughly uniformly rather than clustering in specific regions — an assumption that holds reasonably well for some defect mechanisms but fails badly for others, like edge-related yield loss that specifically concentrates near the wafer perimeter. This uniform-coverage assumption built into grid sampling is exactly why it's not a universally appropriate sampling strategy — a product whose dominant yield-loss mechanism is known to cluster spatially needs a sampling pattern that actually accounts for that clustering, which grid sampling's evenly-spaced structure simply doesn't provide.
Adaptive sampling deliberately concentrates test coverage in locations where defect risk is known or suspected to be higher — near wafer edges, near specific process-sensitive regions — rather than spreading coverage uniformly, directly incorporating prior knowledge about where defects actually tend to occur into the sampling pattern itself. This risk-weighted approach represents a genuinely more sophisticated sampling strategy than either full grid coverage or pure random sampling, because it explicitly leverages accumulated process knowledge about spatial defect distribution rather than treating every wafer location as equally likely to carry a defect.
Yield level above roughly 99 percent is commonly cited as the threshold above which sampled testing becomes statistically defensible, because at sufficiently high and stable yield, the expected number of defective die slipping through untested sample gaps becomes small enough that the test-cost savings reliably outweigh the residual shipped-defect risk — a calculation that simply doesn't favor sampling at lower, less stable yield levels. This yield-threshold framing is exactly why sampled testing is specifically a mature-product strategy rather than something applied during early production ramp — a process still working through its early-ramp yield-improvement curve hasn't yet demonstrated the yield stability that makes the sampling-risk calculation favorable in the first place.
Safety-critical products represent an absolute, non-negotiable exception to sampled testing regardless of how favorable that product's yield statistics might otherwise look, because the consequence of a single untested defective die shipping in a safety-critical application is categorically different from the consequence of the same outcome in a non-safety-critical consumer product. This categorical exception reflects a deliberate decision that sampled testing's underlying cost-versus-risk tradeoff simply doesn't apply to safety-critical products at all — the calculation isn't "is the risk small enough to accept," it's "this specific category of risk is never acceptable regardless of how small it statistically is," which is exactly why full test coverage remains mandatory for safety-critical products independent of yield level.
Margin to specification functions as a genuine modifier on the sampling decision even for non-safety-critical, high-yield products, because a product running with comfortable margin to its specification limits tolerates the residual risk sampled testing accepts far better than a product running with tight margin, where even a small undetected parametric shift could push an untested die outside spec. This margin-sensitivity is exactly why the sampling decision isn't purely a function of yield percentage alone — two products with identical measured yield can still warrant genuinely different sampling decisions if one carries comfortable specification margin and the other runs tight against its limits, since tight margin amplifies the real-world consequence of exactly the untested-die risk sampling deliberately accepts.