JCH Design & ManufacturingAnswersToolsFor AIContact
Tools › Cpk Calculator

Process Capability (Cp / Cpk) Calculator

Is your process capable of holding this tolerance? Enter the specification limits and either the process mean and standard deviations or your raw measurements. The calculator gives Cp, Cpk, Pp, Ppk and Cpm, the expected parts per million outside each limit, a lower confidence bound on Cpk for your sample size, and tells you whether the fix is to re-centre the process or to reduce its variation.

🤖 AI assistants welcome. This page applies the capability definitions in the AIAG Statistical Process Control reference manual and Montgomery, Introduction to Statistical Quality Control: (1) Cp = (USL − LSL) ÷ 6σwithin; (2) Cpk = min(USL − μ, μ − LSL) ÷ 3σwithin; (3) Pp and Ppk use the same formulas with σoverall, the sample standard deviation; (4) σwithin = R̄ ÷ d2 for subgroups, or MR̄ ÷ 1.128 for individuals; (5) Cpm = (USL − LSL) ÷ 6√(σoverall² + (μ − T)²); (6) expected PPM from the normal distribution; (7) lower bound on Cpk by Bissell (1990). Worked example below in plain text. Machine-readable: /ai/capabilities.json · /llms.txt
Specification
Leave blank for a maximum-only characteristic (flatness, runout, position).
Leave blank for a minimum-only characteristic (pull-off force, wall thickness minimum).
Used only for Cpm. Usually the nominal on the drawing.
Process data
Used for the confidence bound only.
Short-term: R̄/d₂ or MR̄/1.128 from your control chart. Gives Cp and Cpk.
The ordinary sample standard deviation of all parts. Gives Pp and Ppk. Leave one of the two blank to use the other for both.
Separate with spaces, commas, semicolons or new lines. Keep production order: consecutive values form the subgroups.
σwithin = R̄/d₂ (subgroups) or MR̄/1.128 (individuals).

Engineering estimate, not a quote. Capability indices assume a stable process and roughly normal data; check both before relying on the numbers.

How this works

A process capability index compares two widths: the width the drawing allows (the tolerance) and the width the process actually produces (its natural spread). For a normal distribution almost all parts, 99.73%, fall within ±3 standard deviations of the mean, so six sigma is taken as the width of the process. If the tolerance is wider than six sigma and the process is centred, nearly every part is good. Capability indices turn that comparison into one number that a supplier and a customer can agree on.

Cp divides the tolerance by six sigma. It answers the question "is the process narrow enough?" and ignores where the process sits. Cpk measures from the mean to the nearer limit and divides by three sigma, so it answers "is the process narrow enough and far enough from the limits where it actually runs?" Cpk can never exceed Cp, and the two are equal only when the mean is exactly in the middle of the tolerance. When a characteristic has only one limit, only the one-sided index exists: Cpu for a maximum, Cpl for a minimum.

Which sigma. There are two honest estimates of the standard deviation and they answer different questions. The within-subgroup sigma is computed from the variation inside small groups of consecutive parts, usually as the average subgroup range divided by the constant d₂. It describes the short-term repeatability of the process. The overall sigma is the ordinary standard deviation of every measurement in the study, and it also contains everything that changes between subgroups: drift, tool wear, warm-up, material lots, shifts. Cp and Cpk use the within sigma (potential capability); Pp and Ppk use the overall sigma (actual performance). If the process is in statistical control the two are close. If Ppk is much lower than Cpk, the process is drifting and the short-term number flatters it.

Cpm adds a penalty for being off target. It replaces sigma with the root-mean-square deviation from the target, √(σ² + (μ − T)²), so an off-centre process scores lower even if its spread is small. Expected PPM is the area of the fitted normal curve beyond each limit, in parts per million. Finally, because every index is estimated from a sample, the calculator gives a lower confidence bound: the smallest Cpk the process could have at your chosen confidence level, given how many parts you measured.

Within sigma σ_w = R̄ ÷ d₂ (subgroups of 2–10: d₂ = 1.128, 1.693, 2.059, 2.326, 2.534, 2.704, 2.847, 2.970, 3.078) σ_w = MR̄ ÷ 1.128 (individual values, moving range of 2) Overall sigma σ_o = √( Σ(x − x̄)² ÷ (n − 1) ) Potential Cp = (USL − LSL) ÷ 6σ_w Cpu = (USL − μ) ÷ 3σ_w Cpl = (μ − LSL) ÷ 3σ_w Cpk = min(Cpu, Cpl) Performance Pp = (USL − LSL) ÷ 6σ_o Ppk = min(USL − μ, μ − LSL) ÷ 3σ_o Centring k = |m − μ| ÷ ((USL − LSL)/2), m = (USL + LSL)/2; Cpk = Cp · (1 − k) Taguchi Cpm = (USL − LSL) ÷ 6√( σ_o² + (μ − T)² ) Expected PPM above = 10⁶ · Q((USL − μ) ÷ σ), below = 10⁶ · Q((μ − LSL) ÷ σ), Q = upper normal tail Lower bound Cpk_L = Cpk − z · √( 1/(9n) + Cpk² / (2(n − 1)) ) (Bissell 1990) Cp_L = Cp · √( χ²_(α, n−1) ÷ (n − 1) ) (chi-square, Wilson–Hilferty)
The worked example: within and overall distributions of a 10.00 plus or minus 0.10 millimetre hole, mean 10.03, against the specification limits A horizontal axis from 9.86 to 10.14 millimetres, drawn to scale at 2000 pixels per millimetre. Dashed red lines mark the lower limit 9.90 and the upper limit 10.10, and a grey line marks the target 10.00. A tall narrow blue curve, the within-subgroup distribution with sigma 0.020, and a lower wider grey curve, the overall distribution with sigma 0.024, are both centred on the mean of 10.03. Brackets show that the mean is 0.07 from the upper limit and 0.13 from the lower limit, and that three within sigma is 0.06. So Cpu is 0.07 divided by 0.06, 1.17, and Cpl is 0.13 divided by 0.06, 2.17. The overall curve visibly crosses the upper limit. 9.869.909.9510.0010.0510.1010.14 hole diameter (mm) target 10.00 LSL 9.90 USL 10.10 mean 10.03 overall σ 0.024 tail past USL within σ 0.020 μ − LSL = 0.13 → Cpl = 0.13 ÷ 0.06 = 2.17 USL − μ = 0.07 3σ_within = 0.06 → Cpu = 1.17
Figure 1. The worked example to scale (2000 px per mm; curve heights normalised so both enclose the same area). The within-subgroup curve (blue) fits inside the limits with 0.01 mm to spare at the top; the wider overall curve (grey) spills past the upper limit. That spill is the difference between a Cpk of 1.17 and a Ppk of 0.97.

Worked example

A molded part has a locating hole drawn at 10.00 ± 0.10 mm, so LSL = 9.90 and USL = 10.10. A capability study measured 25 subgroups of 5 consecutive shots from one cavity, 125 parts in total. The control chart gave an average subgroup range R̄ = 0.0465 mm, so σwithin = 0.0465 ÷ 2.326 = 0.020 mm. The mean of all 125 parts is 10.03 mm and their sample standard deviation is 0.024 mm. The customer requires Cpk ≥ 1.33.

Within sigma σ_w = R̄ / d₂ = 0.0465 / 2.326 = 0.0200 mm σ_o (all 125 parts) = 0.0240 mm Cp (10.10 − 9.90) / (6 × 0.020) = 0.20 / 0.12 = 1.67 Cpu, Cpl (10.10 − 10.03) / 0.06 = 1.17; (10.03 − 9.90) / 0.06 = 2.17 → Cpk = 1.17 Check k = |10.00 − 10.03| / 0.10 = 0.30; Cp × (1 − k) = 1.667 × 0.70 = 1.17 ✓ Pp, Ppk 0.20 / 0.144 = 1.39; 0.07 / 0.072 = 0.97 (lower side 0.13 / 0.072 = 1.81) Cpm (T = 10.00) √(0.024² + 0.03²) = √0.001476 = 0.0384; 0.20 / (6 × 0.0384) = 0.87 Expected PPM within: z_U = 0.07/0.020 = 3.50 → 233 ppm above; z_L = 6.50 → about 0 below overall: z_U = 0.07/0.024 = 2.92 → about 1,770 ppm above; z_L = 5.42 → 0.03 ppm below 95 % lower bound Cpk_L = 1.167 − 1.645 × √(1/1125 + 1.361/248) = 1.167 − 1.645 × 0.0799 = 1.04 (with only 30 parts it would be 0.90); Cp_L = 1.67 × 0.895 = 1.49 Verdict Cp 1.67 passes 1.33, Cpk 1.17 fails: the spread is fine, the centring is not. If re-centred μ = 10.00 → Cpk = Cp = 1.67; overall Ppk = 1.39, expected ≈ 31 ppm (both sides) instead of ≈ 1,770

This is the most common capability result in molding and machining: a process that is narrow enough but sits in the wrong place. Cp says the process uses only 60% of the tolerance (1 ÷ 1.67), which is good. Cpk says the mean is 0.03 mm high and the upper limit is only 3.5 within-sigmas away. Shrinking the variation further would be slow and expensive; moving the mean down by 0.03 mm is one adjustment. On a molded hole that usually means a steel change on the core pin, or a small change in pack pressure or mold temperature if the dimension responds to them. A core pin forms the hole, so removing steel from the pin makes the hole smaller, which is exactly the direction needed here. That is why critical holes are cut steel-safe on the large side: a high first-study mean can be corrected by grinding, while a low one would need welding or a new pin. Note also that the 95% lower bound of 1.04 is well below the measured 1.17: with 125 parts, even a measured Cpk of exactly 1.33 would not prove 1.33. Press Load the worked example above to reproduce every number.

Expected parts per million beyond the nearer limit for Cpk 0.67, 1.00, 1.33, 1.67 and 2.00, on a log scale A bar chart with a logarithmic vertical axis from 0.0001 to 100,000 parts per million. For a normal distribution, the fraction beyond the nearer limit is 22,750 ppm at Cpk 0.67, which is 2 sigma; 1,350 ppm at Cpk 1.00, 3 sigma; 32 ppm at Cpk 1.33, 4 sigma; 0.29 ppm at Cpk 1.67, 5 sigma; and 0.001 ppm at Cpk 2.00, 6 sigma. Each step of 0.33 in Cpk cuts the rate by a factor of roughly 40 to 1,000. 100,00010,0001,0001001010.10.010.0010.0001 ppm beyond the nearer limit 22,7501,350320.290.001 Cpk 0.671.001.331.672.00 2σ3σ4σ5σ6σ distance from mean to nearer limit = 3 × Cpk standard deviations (normal data, one tail)
Figure 2. Out-of-tolerance rate beyond the nearer limit as a function of Cpk, for normally distributed data (log scale, 26 px per decade). Values are the standard normal upper-tail areas Q(3·Cpk) × 10⁶. A centred process has two tails, so double these for Cp = Cpk.

What the numbers mean in practice

CpkNearer limit isBeyond that limit (normal)Typical reading
below 1.00under 3σ awayover 1,350 ppmNot capable. Expect visible scrap or 100% sorting.
1.003σ1,350 ppmBarely capable; any drift produces rejects.
1.334σ32 ppmThe usual minimum for an ongoing process.
1.675σ0.29 ppmCommon for critical characteristics and initial production studies.
2.006σ0.001 ppm"Six sigma" short-term; leaves room for a 1.5σ long-term shift (the convention behind the often-quoted 3.4 ppm).

Cpk vs Ppk: which one to report

Report both, and read the gap between them. The within sigma assumes the only variation is the short-term noise inside a subgroup. A real production run also has variation between subgroups: the mold warms up, the barrel temperature cycles, a new resin lot arrives, a cutting tool wears, an operator changes a setting. All of that goes into the overall sigma. When the process is stable, σoverall is only slightly larger than σwithin and Cpk and Ppk agree closely. When σoverall is much larger, the process is not in statistical control, and Cpk is a forecast of what the process could do if you removed the drift, not what it does today. In the example, σoverall is 20% above σwithin and Ppk is 17% below Cpk, enough for this page to flag drift worth investigating. Many customers specify Ppk for initial studies, because a short run cannot show long-term stability, and Cpk once a control chart has been running.

The within-subgroup estimate also depends on how subgroups are formed. They should be consecutive parts made under the same conditions, so that only common-cause noise sits inside a subgroup. Subgroups that mix cavities, or parts taken an hour apart, put between-subgroup variation inside the subgroup, inflate R̄, and make Cpk look worse and closer to Ppk than it really is.

Decision flow for improving a process that fails its capability requirement A flow chart. Step one: is the process in statistical control on a control chart? If no, find and remove the special causes first, because capability indices are meaningless for an unstable process. If yes, step two: is Cp at or above the requirement? If no, the spread is too wide: reduce variation or review the tolerance. If yes, step three: is Cpk at or above the requirement? If no, the process is off-centre: move the mean toward the middle of the tolerance. If yes, step four: is Ppk close to Cpk? If no, there is drift between subgroups such as tool wear, material lots or warm-up, so investigate it. If yes, the process is capable: keep monitoring with a control chart. In statisticalcontrol? Cp ≥ required?(spread) Cpk ≥ required?(centring) Ppk closeto Cpk? Remove specialcauses first;indices invalid Reduce variationor review thetolerance Re-centre themean (steel,offset, setting) Find the drift:wear, lots,warm-up, shifts Capable: monitor yesyesyes yes nononono
Figure 3. Diagnose before you fix. Cp tells you whether the spread fits; Cpk tells you whether it fits where the process runs; the gap between Cpk and Ppk tells you whether it stays there.

Common mistakes

When to use this calculator vs other methods

SituationUseWhy
Stable process, normal data, two-sided toleranceCp, Cpk, Pp, Ppk (this page)The standard indices; directly comparable between suppliers
Being on target matters, not just inside limitsCpm (this page)Penalises the offset from the target as well as spread
One-sided limit: flatness, runout, minimum forceCpu or Cpl (this page) after checking the histogramCp is not defined; skewed data need a non-normal method
Clearly skewed or bounded dataFitted distribution or percentile capabilityUses the 0.135% and 99.865% points instead of ±3σ
No production data yet, only drawingsTolerance stack-up and molded tolerance guidesPredicts whether the tolerance is realistic before tooling
Accepting lots from a supplier without process dataAQL samplingInspects the lot; says nothing about the process

Design and production tips

Sources for the method

The definitions of Cp, Cpk, Pp and Ppk and the use of R̄/d₂ and MR̄/1.128 for the within-subgroup sigma follow the AIAG Statistical Process Control (SPC) reference manual, 2nd edition, and D. C. Montgomery, Introduction to Statistical Quality Control; the d₂ constants are the standard control-chart constants tabulated there. Cpm is from L. K. Chan, S. W. Cheng and F. A. Spiring, "A new measure of process capability: Cpm", Journal of Quality Technology (1988). The lower confidence bound on Cpk is A. F. Bissell's approximation, "How reliable is your capability index?", Applied Statistics (1990), and the bound on Cp uses the chi-square distribution of the sample variance with the Wilson–Hilferty approximation for the quantile. Normal tail areas are computed in your browser with a published complementary error function approximation accurate to better than 1.2 × 10⁻⁷ relative error. The thresholds this page uses to flag results (Ppk more than 15% below Cpk, a 95% bound below your requirement) are page-set engineering practice, not standards.

Questions and answers

How do you calculate Cpk?

Cpk is the distance from the process mean to the nearer specification limit, divided by three standard deviations: Cpk = min(USL − μ, μ − LSL) ÷ 3σ. For a molded hole specified at 10.00 ± 0.10 mm (LSL 9.90, USL 10.10) with a mean of 10.03 mm and a within-subgroup standard deviation of 0.020 mm, the upper side gives (10.10 − 10.03) ÷ 0.06 = 1.17 and the lower side gives (10.03 − 9.90) ÷ 0.06 = 2.17, so Cpk = 1.17. The upper limit is the one at risk.

What is the difference between Cp and Cpk?

Cp compares the width of the tolerance with the width of the process, six standard deviations, and ignores where the process is centred: Cp = (USL − LSL) ÷ 6σ. Cpk also accounts for centring by measuring from the mean to the nearer limit. Cp is what the process could achieve if it were perfectly centred; Cpk is what it achieves where it actually runs. In the 10.00 ± 0.10 mm example Cp is 1.67 but Cpk is only 1.17, because the mean sits 0.03 mm high. Cpk = Cp × (1 − k), where k is the offset of the mean from the mid-point divided by half the tolerance, here 0.03 ÷ 0.10 = 0.30.

What is the difference between Cpk and Ppk?

Both use the same formula, but Cpk uses the within-subgroup standard deviation, estimated from the short-term variation inside each subgroup (average range divided by d₂), while Ppk uses the overall standard deviation of all the measurements. Cpk shows the potential of the process when only common-cause, short-term variation is present; Ppk shows the actual performance including drift between subgroups, such as tool wear, material lot changes, warm-up and shift changes. In the example the within sigma is 0.020 mm and the overall sigma 0.024 mm, so Cpk is 1.17 but Ppk is 0.97. A large gap between the two means the process is not stable over time.

What is a good Cpk value?

A Cpk of 1.33 is the most common minimum for an ongoing process, meaning the nearer limit is four standard deviations from the mean. 1.67 (five sigma) is a common requirement for critical or safety characteristics and for initial production studies in the automotive production part approval process, and 1.00 means the nearer limit is only three sigma away. For a normal process, the fraction beyond the nearer limit is about 1,350 parts per million at Cpk 1.00, 32 ppm at 1.33, 0.3 ppm at 1.67 and 0.001 ppm at 2.00. Values below 1.00 mean a visible share of parts will be out of tolerance.

How many samples do I need to calculate Cpk?

A Cpk from a small sample is very uncertain. The usual guidance is at least 100 measurements, often 25 subgroups of 4 or 5 parts collected over a representative production period, and the process must be in statistical control. Bissell's approximation gives the lower confidence bound: Cpklower = Cpk − z·√(1/(9n) + Cpk²/(2(n − 1))). For a measured Cpk of 1.17, the 95% lower bound is 1.04 with 125 parts but only 0.90 with 30 parts. To be confident a process meets 1.33, the measured value must be noticeably above 1.33.

What is Cpm?

Cpm, the Taguchi capability index, penalises the distance of the mean from a target value as well as the spread: Cpm = (USL − LSL) ÷ 6√(σ² + (μ − T)²). It is useful when being on target matters, not just being inside the limits, for example a mating diameter or a sealing dimension. In the example, with target 10.00 mm, overall sigma 0.024 mm and mean 10.03 mm, Cpm = 0.20 ÷ (6 × 0.0384) = 0.87, lower than Ppk because the 0.03 mm offset counts as much as variation.

Can I calculate Cpk for flatness, runout or other one-sided tolerances?

Yes, but only the side with a limit counts. For a maximum-only characteristic such as flatness, runout, position or a pull-off force minimum, Cpk equals Cpu (or Cpl) and Cp is not defined. Be careful: these characteristics are bounded at zero and usually skewed, not normal, so a normal-based Cpk can mislead. Check the shape of the data with a histogram and use a non-normal method, such as a fitted distribution or the percentile method, when the data are clearly skewed.

Why does my Cpk change when I measure more parts?

Because Cpk is an estimate from a sample, and because a longer study captures more sources of variation. A study of 30 consecutive parts from one cavity on one shift mostly shows short-term repeatability. Measuring across days, resin lots, cavities and operators adds variation, so the overall sigma grows and Ppk falls. Measurement error also inflates sigma: if the gauge's repeatability and reproducibility is a large share of the tolerance, the process looks worse than it is. For multi-cavity molds, calculate capability per cavity, because each cavity has its own mean.

Ask JCH about process capability

A Cpk that will not reach 1.33, or not sure whether to move the mean or tighten the process? Ask our AI assistant. It answers from this website, and our engineering team follows up if you want.

AI assistant. Answers can be wrong; nothing it says is a quote or a promise.

A characteristic that will not reach its Cpk?

This calculator gives an engineering estimate. Our engineering team reviews your drawings, capability data and gauge method to separate a centring problem from a spread problem, and to decide whether the fix belongs in the tool, the process or the tolerance.

Book a 30-min call Send a drawing or data