Key takeaways
  • Cp compares process spread against the specification width. Cpk also accounts for how far off-centre the process sits.
  • Cpk can never exceed Cp. A large gap between them means the process is capable but badly centred, which is usually the easier fix.
  • Capability calculated on an unstable process describes that particular week, not the process.
Cp versus Cpk in one line
Cp asks if it could fit; Cpk asks if it does fit

A process with tight variation sitting near one limit has a good Cp and a poor Cpk. That gap is good news: centring a process is normally far easier than reducing its variation.

The two indices

  • Cp = (upper spec minus lower spec) divided by six standard deviations. It compares the width available against the width used, ignoring position.
  • Cpk = the smaller of (upper spec minus mean) and (mean minus lower spec), each divided by three standard deviations. It penalises being off-centre.

What the numbers mean

  • Cpk below 1.0: the process is producing out-of-specification output as a matter of course.
  • Cpk of 1.00: the spread just fits the specification, with no margin for drift.
  • Cpk of 1.33: a widely used minimum for ongoing production.
  • Cpk of 1.67 and above: often required for safety-related or critical characteristics.
  • Customer requirements vary; automotive and aerospace commonly specify their own thresholds.

Stability first

Capability assumes the process is stable, because it projects the observed spread forward. If the process is drifting or shifting, the standard deviation you calculated belongs to a mixture of processes and the resulting index predicts nothing. Establish control with a control chart, then calculate capability.

Cp, Cpk, Pp, Ppk

Pp and Ppk use the overall standard deviation of all data, while Cp and Cpk use the within-subgroup variation. Ppk therefore includes drift between subgroups and is usually the lower, more honest number. A large gap between Cpk and Ppk means the process moves between subgroups even if each subgroup looks tight.

When capability is the wrong tool

Capability indices assume roughly normal data and a two-sided specification. For skewed distributions, one-sided limits, or attribute data such as pass and fail, they mislead. Percent out of specification, or a defect rate, is the more honest measure in those cases.