Cp, Cpk, Pp, and Ppk are variables-data metrics — they need a continuous measurement, a mean, and a standard deviation to calculate. A meaningful share of real processes don't produce that kind of data at every step: pass/fail functional tests, visual inspections, multi-step assemblies where the meaningful output at each stage is simply "did this step complete without a defect." Those processes still have capability worth quantifying — it just isn't expressed as a Cpk, and treating attribute-based process performance as if it needs to be forced into a Cpk-shaped answer usually means either the wrong metric gets reported or a genuinely different, and often more useful, set of metrics gets skipped entirely.
Defects Per Unit
The starting metric for attribute-based process performance is defects per unit (DPU) — total defects found divided by total units produced, over some meaningful measurement window. It's a direct, easily understood rate, and it forms the basis for the metrics built on top of it.
DPU on its own has a limitation worth noting: it treats a unit with three defects the same as three separate units each with one defect, in terms of the total count, which is appropriate for tracking overall defect volume but doesn't distinguish "many units mostly fine, a few bad ones with multiple issues" from "moderate, evenly distributed problems across most units" — two process behaviors that DPU alone reports identically but that usually call for different corrective approaches.
Rolled Throughput Yield
For a multi-step process — an assembly line with several sequential operations, each with its own opportunity to introduce a defect — first-pass yield at any single station only tells part of the story. Rolled throughput yield (RTY) multiplies the first-pass yield of every step together, giving the probability that a unit makes it through the entire process defect-free on the first attempt, with no rework at any station.
RTY is often dramatically lower than any individual step's yield would suggest, and that gap is the whole point of calculating it. A ten-step process where every single station runs a comfortable-looking 98% first-pass yield produces a rolled throughput yield around 82% — meaning nearly one in five units needs rework or repair somewhere along the line, a very different picture than "every station looks fine" suggests when each step is evaluated in isolation. RTY is what exposes that gap, and it's a common blind spot in processes where individual station metrics all look acceptable while the overall process performance quietly isn't.
Z-Bench: The Attribute Equivalent of Cpk
Z-bench (sometimes called Z-score or sigma value in this context) translates a DPU or overall yield figure into an equivalent normal-distribution Z-score — effectively answering "if this defect rate were coming from a normal distribution against a spec limit the way Cpk assumes, how far would the mean be from that limit, in standard deviations." This gives attribute-based processes a number that's directly comparable, in a rough sense, to the sigma-level language used for variables-based capability, without requiring the underlying data to actually be continuous or normally distributed in the first place.
The comparison carries the same caveat that applies to sigma-level conversions generally: Z-bench is a translation for communication purposes, not a literal statement that the attribute process follows a normal distribution. It's useful specifically because it lets a multi-step assembly process and a continuous machining process get discussed on roughly the same capability scale, even though the underlying data and calculation methods are completely different.
Where Teams Misapply This
The most common mistake is forcing genuinely continuous data into an attribute-style DPU/RTY framework out of habit — collapsing a measurable dimension into pass/fail and reporting DPU, when the underlying measurement would support a full Cpk calculation with far more statistical power, the same information-loss problem that applies to downgrading any continuous characteristic to attribute data unnecessarily.
The second common mistake runs the other direction: trying to back-calculate a Cpk-equivalent number for a process that's genuinely attribute-based — no underlying continuous measurement exists at all, just pass/fail outcomes at each station — using Z-bench as if it were interchangeable with an actual Cpk rather than a communication translation. Z-bench is a legitimate and useful number in its own right; it isn't a substitute for the more statistically grounded capability picture that only exists when real measurement data is available.
Calculating Both Kinds of Capability
For processes with genuine continuous measurement data, SigmaDesk's process capability calculator calculates Cp, Cpk, Pp, and Ppk with a built-in normality check, free in the browser, as part of the full SigmaDesk SPC platform alongside control charts and attribute charts for the pass/fail data DPU and RTY are built from.
Not every process capability question has a Cpk-shaped answer. DPU and rolled throughput yield exist precisely for the processes where it doesn't — and they're worth calculating on their own terms, not as a workaround for missing continuous data.