Defects per unit (DPU) is the average number of defects found per unit produced: DPU = total defects ÷ total units. Unlike a pass/fail defective count, it counts every flaw, so a unit with three defects counts as three, capturing the total defect load that a yield number quietly hides.
The distinction DPU insists on, defects, not defectives, is the whole point, and it is the one most plants blur. A first-pass yield says a unit passed or failed; it cannot tell the unit that scraped by with one flaw from the unit riddled with five. DPU can, and that extra resolution is what makes it a better early-warning signal for a process drifting toward trouble. This guide defines DPU, draws the line between defects and defectives, works the calculation, and shows the clean mathematical bridge from DPU to yield. It sits next to defects per million opportunities and first-pass yield in the family of quality counts.
What is defects per unit?
DPU is total defects divided by total units inspected over a period. If 500 units yield 65 defects, DPU is 0.13, on average, thirteen defects for every hundred units. The number can exceed 1.0, and that is a feature, not an error: a DPU of 1.4 means the average unit carries more than one defect, which a yield figure capped at 100% can never reveal. DPU measures the amount of defectiveness, not merely its presence.
That makes DPU a load metric. Where first-pass yield answers “what fraction of units were clean?” DPU answers “how much rework and scrap is really in there?” Two lines can post the same 92% first-pass yield while one carries a DPU of 0.08 and the other 0.35, the second is hiding a much larger defect load inside its 8% of failing units, and only DPU shows it. It is the metric that catches the unit with three problems, not just the fact that a unit failed.
Why does the defect-versus-defective distinction matter?
It matters because the two counts diverge exactly when a process is getting worse, and the divergence is the early warning. A defective is a unit with one or more defects, a binary verdict. A defect is a single nonconformity, and one unit can hold several. When a process starts to slip, units do not fail one at a time in tidy increments; they start accumulating multiple flaws each. A defectives count, and any yield built on it, saturates: once a unit is failed, a second and third defect on it change nothing. DPU keeps counting, so it registers the deterioration a shift or two before the yield number moves enough to alarm anyone.
The practical consequence is where you send attention. If you track only defectives, a unit with three problems and a unit with one look identical in the data, and your root-cause work loses the strongest clue it had, the units piling up multiple, possibly related defects. DPU preserves that clue. It is also why DPU pairs naturally with a Pareto of defect types: the total defect count it is built from is exactly what a Pareto needs.
How do you calculate defects per unit?
The calculation is short, but each step has a definition that has to hold steady:
- Define a defect operationally. Write down what counts as one nonconformity so two inspectors tally the same event identically. Ambiguous definitions make DPU drift for reasons that have nothing to do with the process.
- Count total defects. Add up every nonconformity across the sample, three flaws on one unit is three, not one. This is the count that separates DPU from a defectives rate.
- Count total units. Every unit inspected in the period, defective or not. Keep the sample boundary fixed so the ratio is stable.
- Divide. DPU = total defects ÷ total units. Values above 1.0 are valid and meaningful, they say the average unit carries more than one defect.
- Track by defect type. Break the total into a Pareto of defect categories so DPU points at what to fix, not just how much there is. The total is only a thermometer; the breakdown is the diagnosis.
How does DPU relate to yield?
DPU converts to yield through a clean piece of Poisson math, which is what ties the whole Six Sigma metric family together. If defects occur randomly and independently, the probability a unit has zero defects, its throughput (first-time) yield, is:
First-time yield ≈ e−DPU
So a step with DPU 0.10 has a throughput yield of about e−0.10 = 90.5%; a step at DPU 0.30 yields about 74.1%. Run several steps in series and multiply their yields, and you get rolled throughput yield equivalently, e raised to the negative sum of the step DPUs. The relationship also runs backward: total DPU ≈ −ln(rolled throughput yield). This is why you cannot judge a multi-step process by its final inspection alone: each step leaks a little yield, and DPU is the currency that adds up across them.
That additivity is the practical gift. Final inspection shows one number for the whole line and hides which step is bleeding; the per-step DPUs sum to the total and point straight at the worst offender. A four-step line where every step looks respectable can still lose a quarter of its units to the “hidden factory” of rework between steps, and only step-level DPU makes that visible.
How is DPU different from DPMO and first-pass yield?
All three count quality, but they normalize differently, and choosing wrong sends you chasing the wrong process:
| Metric | What it divides by | Best for |
|---|---|---|
| Defects per unit (DPU) | Units | Sizing total defect load; feeding a Pareto; bridging to yield |
| DPMO | Opportunities (units × opps/unit) | Comparing products of different complexity on equal terms |
| First-pass yield | Units (pass/fail) | Simple gate: what share came out clean the first time |
Read together they are powerful. DPU sizes the load, DPMO makes it comparable across complexity, and first-pass yield gives the headline gate everyone understands. The one trap is reading first-pass yield alone on a complex product and concluding the process is fine because most units passed, DPU on the failing units often tells a harsher story. They all draw from the same defect-tracking record, which is why counting each defect at the source, as it happens, is what keeps every one of them honest.
How do you use DPU to improve?
Use DPU as the thermometer and its Pareto breakdown as the map. Because DPU is a rate, it is fair to compare across time and across lines running the same product, so it works well as a tracked line on a KPI board. Watch the trend, and when it climbs, go to the defect-type Pareto to see which category is driving it. Take the top category, not a random defect, into root-cause work, install a control, and confirm the DPU line comes back down. For a stable process, plot DPU on a control chart so you can tell a real shift from ordinary noise, the same discipline as statistical process control.
The reason DPU deserves real-time collection is the same reason it beats a defectives count: the multi-defect units that warn you earliest are exactly the ones a memory-based end-of-shift tally rounds away. Counting defect by defect at the point of inspection keeps the early-warning signal intact (see the platform). Every point of DPU you remove is scrap and rework you stop paying for, which ties straight back to the six big losses and the plant’s cost of quality. See how one plant sharpened its quality signal in the CLS case study.
Data & sources
DPU and its Poisson bridge to yield are standard Six Sigma definitions from the quality profession.
- ASQ, the American Society for Quality, documents the Six Sigma metric chain, DPU, DPMO, and rolled throughput yield, including the yield relationships; see its rolled throughput yield and Six Sigma resources.
- The throughput-yield relationship, first-time yield ≈ e−DPU follows from modeling defects as a Poisson process; equivalently, total DPU ≈ −ln(rolled throughput yield), the entry point that leads DPU to DPMO to a sigma level.
- ISO 22400-2, the international standard for manufacturing KPIs defines the quality-ratio indicators DPU supports, so plant reporting stays aligned to a recognized definition rather than a house metric.