The Day the CPK Report Came Out, the Whole Room Was Silent for Three Seconds, and I Only Said One Sentence
I still remember two years ago, the yield of that new process wafer was disastrously low. The boss's face turned ashen, and he called together heads from process, equipment, and quality assurance for a "review meeting." The atmosphere in the conference room was heavy. When Xiao Chen from the QA department showed the CPK report, the whole room was indeed silent for three seconds. CPK 0.85, an incredibly low level. The boss pointed directly at me: "Old Wang, tell me, how are we going to fix this?" I took a deep breath, looked at the dense data on the report, and a thought flashed through my mind: "This isn't a CPK problem at all; it's that we've been using the wrong sampling standard!"
What's the Problem? It's Not That CPK Isn't Good Enough, Your Sampling Is Fundamentally Wrong!
To be frank, many times when we evaluate product quality, especially in sampling inspection, we default to using the ANSI/ASQ Z1.4 standard, which is also commonly referred to as the commercial version of MIL-STD-105E. This standard is classic and very useful, but it has a core assumption: your product defect rate is stable. It primarily looks at the "number of nonconforming items," meaning how many NGs you find in your sample. You tell me, in our semiconductor fabs, where DPMO can easily be hundreds or thousands, where are you going to find so many "nonconforming items" to count? What we care more about is "variation"!
This is where MIL-STD-1916 comes into play. The fundamental difference between it and Z1.4 is that it places greater emphasis on variables sampling. What does that mean? Z1.4 is like buying fruit at a market, grabbing a few casually to see if any are rotten. What about 1916? It's like performing precise measurements in a lab, measuring the sugar content, acidity, and size of each piece of fruit, and then analyzing the fluctuations in these data. In other words, Z1.4 counts "nonconforming items," while 1916 measures the "degree of goodness or badness."
How to Do It in Practice? It Depends on Whether You're Counting NGs or Measuring Variation
Frankly speaking, if your process is already mature, with a stable and relatively high defect rate—for example, a DPMO around 10,000, which translates to a 99% yield—then Z1.4's attributes sampling (counting type) will be more intuitive. You set an AQL (Acceptable Quality Level), say 1%, sample 100 units, and if more than 2 units are NG, then the batch is deemed problematic.
However, if your process requirements are very high and the defect rate is extremely low—for example, we often encounter DPMO 6210 (CPK 1.08) or even lower—and you still use Z1.4 to count nonconforming items, you might find that you sample 1000 units and still haven't seen a single NG, making it impossible to make any judgment. This is where 1916's variables sampling comes in handy. It doesn't just look at whether there are NGs; it places more importance on your process capability indices, such as CPK and PPK. It will require you to measure certain key parameters of the product (e.g., voltage, current, thickness), then calculate the mean and standard deviation of these measured values, and based on these, judge the quality of the entire batch. Therefore, the key is that 1916 is about assessing the stability and consistency of the process, rather than simply looking for defective products.
The Most Common Pitfall: Good-Looking Numbers Don't Necessarily Mean a Stable Process
To put it plainly, we used to often make a mistake: in order to make the CPK report look good, we might only adjust a specific parameter, pushing the numbers higher, but the overall process stability wouldn't actually improve. With Z1.4's attributes sampling, you might see a decent yield, but behind it, significant variation could be lurking. You'll only discover the problem when you actually encounter products at the edge of the specifications.
I once encountered a case where a certain process's CPK report consistently stayed above 1.33, looking very good. However, clients later complained that our products occasionally experienced "unexpected" crashes on their machines. Tracing the root cause, it was found that although the products were within specifications, the variability of a certain electrical parameter was very large, with a high proportion of products near the specification limits. Using Z1.4 sampling, these "in-spec" products would all be classified as good, making it impossible to detect the problem. If we had implemented 1916 then, enforcing variables sampling to analyze the normal distribution and variation of these parameters, we might have discovered this potential risk earlier.
One Thing You Can Do Today
Re-examine your current process sampling standards and ask yourself: Am I trying to count NGs, or am I trying to measure variation?