Couldn't understand corner simulations.

[sneha rayala]

Hi everybody , I recently entered the workforce and began working on genuine PDKs for the first time. Thus, I was unable to completely understand the process corners in TSMC, such as the numerous corners like FSF, SSF, etc. and FSG, SSG, etc. and monte carlo simulations

Would you kindly point me in the direction of a complete article or a link where a detailed explanation is available? I want to study this topic thoroughly and have a thorough knowledge.
Many thanks.

 

[dick_freebird]

Never seen a "ternary corner" before. Hope to make it a
good while longer.

"Corners" are a fiction of convenience aimed at digital
design closure. They should be barely faster / slower,
stronger / weaker than what the foundry will say is
shippable.

What that means, is highly variable company to company,
technology to technology, "vintage" to "vintage". What
the third "corner" designator is in this case, no idea ("G"?).

One way to approach "corners" development is to do the
mismatch and process stats based on big volume data,
then run testbenches, then scrub the output data for the
best and worst iterations that:

- do not violate any rejectable WAT criterion
- produce the outlier, high or low, for any output param
- encompass the sandbagging your foundry*engineering
management will negotiate (spoiler: your engineering
management lacks balls, and you will be made to suffer)

Now at the end of this you'd have "N" corners which result
in something you should care about. Note this is more of
an "analog" approach; in the more distant past corners
were simply stuff like process WAT limit (plus sand):

"min VT, max kp / min Tox, max delta-L"

permuted across N, P (and flavors) to get FF. FS. SF. SS.

Now people migt argue (correctly) that thie is unreasonable
as some process attributes are shared (like Tox and to some
extent delta-L) and it's unlikely to get truly process-max
TOXN and process-min TOXP at once (to say the least). But
if you used a righteous set of MC statistics to do it, that should
all fall out of the standard deviation and correlation coefficients.

Which you would be wise, to check work. Because CAD weasels
don't have to live with their lies and shortcuts directly, and get
their ticket punched by getting "done" on schedule. To which
additional point, "what constitutes "done"?