That Day, When the CPK Report Came Out, the Entire Room Fell Silent for Three Seconds
I still remember several years ago, a batch of new products in our factory caused a lot of headaches. At that time, the yield was consistently stuck, and no matter how we looked at it, the process seemed fine, but the WAT (Wafer Electrical Test) was always an issue. The boss monitored us daily, putting immense pressure on us. Once, during a review meeting, the CPK report came out: 0.87! The entire room truly fell silent for three seconds; the air instantly froze. Everyone exchanged glances, unsure where the problem actually lay. But frankly, the problem wasn't with the process; it was with how we looked at the data.
Where Did the Problem Lie?
To put it simply, Wafer Electrical Test (WAT) is our 'health check' for each batch of wafers. We measure several key parameters, such as transistor switching speed, leakage current, and so on. If these data points are all within specifications, theoretically, the product should be fine. However, many people only look at whether a single parameter is 'Pass/Fail,' overlooking the 'distribution' behind these parameters. It's like a doctor examining your blood pressure; besides checking if it's above the standard, they also look at the 'trend' and 'stability' of your blood pressure. In WAT, what we need to look at is also trends and stability.
Therefore, the key is that we shouldn't just look at the data for a single wafer, but rather the data distribution for the 'entire batch' of wafers. Even if all parameters pass, if their distribution is so wide that it nearly touches the specification limits, then the stability of that wafer's manufacturing process becomes questionable.
How Is It Done in Practice?
We typically use statistical tools to analyze these WAT data. The most basic ones are CPK (Process Capability Index) and DPMO (Defects Per Million Opportunities).
- Analyzing CPK: CPK is an indicator that measures process capability and stability. Simply put, the higher the CPK value, the more stable your process, and the more concentrated your output products are in the center of the specifications. We generally require a CPK of at least 1.33 to be considered stable. If, as I mentioned earlier, the CPK is only 0.87, it means your process variation is too large, product quality is not stable enough, and the yield naturally won't improve. You can imagine product parameters as a bell curve, and the specification limits as the walls on both sides of the bell. A high CPK means the bell curve is narrow and far from the walls; a low CPK means the bell is wide and already close to touching the walls.
- Analyzing DPMO: This indicator is more intuitive, representing how many defects occur per million opportunities. Suppose we have a parameter specified as 100±10 Ohms. If a batch of wafers is measured, and the average value is 100 Ohms, but 6210 data points have resistance values falling outside 110 Ohms or 90 Ohms, then the DPMO for this batch is 6210. The lower this value, the better, as it directly reflects your defect rate.
In other words, when you see a CPK below 1.33 or a DPMO higher than an acceptable standard (such as 3.4 DPMO), it indicates a potential process issue. At this point, the response should not be to directly adjust process parameters, but rather to conduct a deep analysis of which aspect of the distribution is problematic.
The Most Common Pitfalls
The most common pitfall I've encountered is engineers only looking at the average value. When I first joined the company, I did the same. There was a parameter whose average looked very good, right in the middle of the specifications. However, one time, a senior engineer from the process team pulled out the data and plotted a histogram, only to discover that while the average was fine, the distribution was 'bimodal'! This meant that half of the products were running towards the upper specification limit, and the other half towards the lower specification limit, simply averaging out to the middle. In such a situation, the yield absolutely won't be good, and if the process drifts even slightly, a whole batch will fail immediately. Therefore, simply looking at the average value is far from enough; one must always examine the distribution plot.
Another common mistake is rushing to adjust the process as soon as a poor CPK is observed. Frankly, many times a deteriorating CPK might just be due to increased measurement system error, or a few abnormal data points from a small number of wafers widening the overall distribution. Therefore, it's crucial to first confirm if there's a measurement system issue or to exclude outliers before drawing conclusions.
One Thing You Can Do Today
Next time you report WAT data, don't just look at Pass/Fail; also include CPK and DPMO.