Die shrink qualifications and management

[ewinter]

Memory manufacturing process (sram and dram, etc) regularly go through changes. These changes are typically die shrink. A die shrink yields changes in performance such as faster access time, reduced currents faster rise times, etc. In my past experience it was a normal process to qualify new die revisions through regression testing including design margin testing. Once tested and qualified it was standard practice to create a new part number to define the new die revision part as a unique entity so it would not be mixed with other die revisions. We never mixed die revisions under the same part number.
Can someone enlighten me on process they have used and some of the considerations they may have.

 

[IRstuff]

Sounds like a classic configuration management and control problem. Why would anyone NOT assign new rev letters as a minimum?

You could waste a lot of time and wafers in fab if the incorrect older mask set was pulled up on a die shrunk run.

Atthe finished part level, you need, again, at least a rev letter change, since you should be selling the new part for more money, since it has better speed.

If you sell it under the existing part number, you'd be hard pressed to make back the investment in the mask charges and regression testing incurred.

TTFN

 

[ewinter]

To clarify, I am talking about this from an OEM perspective. We use memory devices in our product and have to control the use of these devices when changes from the manufacturers of memory devices occur. When changes occur I have to be concerned with what the affect of the performance differences will have on the functionality of our product. For that reason and past experience, testing is obviously required and new internal part numbers need to be defined to insure that no die rev "C" devices will be mixed with die rev "D" devices because of the possible performance/functionally impact.
Your thoughts?

 

[IRstuff]

I agree that new part numbers should be employed in the general case. Die shrinks do not always result in uniformly better performance.

If your designers took advantage of obscure or undocumented behavior, mixing parts could seriously affect your overall performance and yield, particularly at temperature extremes.

Your designers ought to re-run worst/best-case circuit simulations with the changed parameters to determine if the entire operating envelope is truly available. This should be done in conjunction with real testing, since it's unlikely that you'll get a sufficient variance in parameters for the new devices to adequately stress your design.

TTFN

 

[designchain]

ewinter,

It's situation-dependent. I agree with IRstuff that the new devices must be simulated and tested in new designs. The practical issues in assigning a new internal part number include

a) whether the new die rev can be positively identified by packaging information - did the mfr rev their part? Can you trust a date-code-based switch-over?

b) how many BOMs get affected and rev'd by what would otherwise be simply another source listed on the AML? How much work is issuing ECOs for each of them every time one of the parts on the AML is revised? How will you manage revisions from each of the manufacturers on the AML? Will you assign them different part numbers, thereby making them all single-sourced? What risks will you run then? How will you manage the resulting multiple internal part numbers for a single board position?

If you can PROVE that the new part is interchangeable with the old through simulation and testing, then why rev or change the part number?

Mike

--
Mike Kirschner
Design Chain Associates, LLC

http://www.designchainassociates.com