What's WAT? An overview of WAT/PCM data
)
Data generation
Wafer Acceptance Testing (WAT) also known as Process Control Monitoring (PCM) data is data generated by the Fab at the end of manufacturing and generally made available to the fabless customer for every wafer. The data will typically have between forty and one hundred tests, each test having a result for each site (or “drop-in”) on the wafer. The sites are located so that the Fab can monitor the consistency of key parameters across the wafer. Within the Fab itself, there is a statistical process control system which uses this (and other) data to improve yield and reduce defects.
Typically the data is sent per lot, so 25 wafers are represented in a single data file. The formats are often ASCII CSV (comma separated value) or Excel. Any modern yield management system (YMS) or yield analysis system will be able to support this type of data.
The fabless model and manufacturing
The high costs of building, resourcing and operating a foundry fabricating integrated circuits are well known. Fabless companies avoid this capital cost and focus on design and innovation in their area of expertise.
On the other hand, the fabless company relies on the expertise and skills of the foundry to produce quality wafers. Many times a process used by a fabless company to manufacture their wafers is shared with many other companies. The foundry will have tests for monitoring quality and controlling their process at every stage of manufacture. The data at the end of the line when the wafers are ready to be shipped is made available to the fabless customer. This relatively small but powerful data is the subject of this article.
Data transfer from the fab
Typically Fabs are located in Asia but some are still located in Europe and the USA. Irrespective of where they are located, the WAT data has often to be transferred for processing into the YMS. YMS vendors should provide a secure method (script) for data transfer from the Fab to their system. YMS these days are typically in the cloud or alternatively “On Premise”. A majority of fabless companies prefer to outsource YMS especially if they start to have reasonable volume and the vendor should be able to provide both cloud and on premise options to the customer.
Processing the data
When the WAT data arrives in the YMS incoming directory it is typically split by wafer id. The data for each wafer is then split into each site (or location of testing on the wafer). The number of sites can vary from five to nine and sometimes much higher in speciality chips. Each site then will have typically sixty results. These results vary from “critical” parameters to “monitoring” parameters. Many tests, especially the critical ones, will have limits. So all this data is fed into the YMS and available for analysis.
)
A good YMS system will process this data automatically into the database in seconds. There should be no need to manually process the data. The data size is nowhere near the size of the STDF (or equivalent) data from wafer sort and final test. In fact, it’s far less than 1% of the size of a YMS database. For example the WAT data used in this article has 62 parameters across nine sites and amounts to 75KB compressed. The associated wafer sort data for the same lot amounts to 1GB compressed. So WAT data is overwhelmingly well-behaved (unlike some of the wafers of course!), of small size and lending itself to fast analysis and visualization.
Comparing the data vs other manufacturing stages (e.g. STDF)
• It’s very “small” data
• There are no test numbers
• Limits and format are not under control of fabless company
• The results are not part of a device’s specification
• The data is by process and not by a customer’s product design
• The format is often (human readable) Excel or CSV versus STDF (human unreadable) or other binary formats which are more suitable to high volume testing
Visualising WAT data
)
In fig 2, you can see WAT data for a full lot for test Isat_1P3. As you can see the same sites from each wafer are connected together using a trend line of the same colour. Wafer 12 site 5 (coloured blue) in this data set is appreciably different from the other data. Also wafer 1 has a greater range for this Isat test than other wafers except for wafer 12.
Another visualisation is fig 3, which shows a box plot over 40 lots of wafers. The highlighted lot demands more analysis and a wafer or two could cause trouble in due course.
)
Fig 3. A WAT parameter trend across all 40 lots (1000 wafers).
Being able to set alerts when the WAT data arrives can save a lot of pain later on. The limits in the data from fab will be for internal use by the fab. Using a yield management platform like yieldHUB, you can set your own effective SPC limits so that you can spot lot-ids and wafers (like the one highlighted in Fig 3) before they cause yield loss or reliability issues later on.
Correlating WAT data against wafer sort and final test
Once you have a means for connecting data from the same wafers across the supply chain, then further power is unleashed from your YMS. Even better if the connection (or genealogy link) is automated – then you have the potential to really see correlations quickly and without any manual preparation. Fig 4 is an example of the mean of a WAT parameter (Y axis) vs the fail rate of a bin in wafer sort (x-axis). In this example the limits on the fab parameter are well outside this chart range. Based on this data, it should be easy to set an alert limit on the YMS system that would identify wafers from the fab data that would later risk high bin 14 in wafer sort. Equally important, this is key information to feed back to the fab and to understand the underlying root cause which could also of course be design related. This is an example of how important the WAT data is, how important it’s visualisation is and how important is the link-up with test data from wafer sort and final test. In best-in-class YMS, the links are completely automated.
)
The same correlations can be accomplished with final test data. A pre-requisite to the accuracy of such correlation charts in final test is that the data in final test is consolidated, ie, that there is one data point per die in final test and that the data is from the last re-screen of each die. With that “cleansing” in place the WAT to final test correlations are “gold”.
Other analysis techniques
There are other ways of analysing WAT data beyond the scope of this introduction. For example, the results for a WAT site can be correlated with the parametric and bin performance of that part of the wafer in wafer sort.
In summary, if you can get your hands on the WAT data and link it to wafer sort and final test data, you should. You will learn a lot more about the root cause of yield loss, reliability and quality. Being able to present reports that correlate across manufacturing to the fab will save you significant costs in the long run.