hi, I'm looking at modelling of the process variations in mos transistor at say 90nm and below. I'm new to this group, and not sure where exactly to post it.
Basically, the foundry guys give the standard deviation data on vt, transconductance and Idss. With this I would like to figure out how the (say)inverter delays are affected. (Essentially try and figure out the distribution).
In my opinion, all the three are not totally uncorrelated to each other. So, I presume selecting two of these should do a good job. Is this on the right track??
And if someone has worked on these things can he/she share some basic ideas with me??
hi Kevin, thanks for the information. But I'm actually looking for something slightly different. Its my mistake, since I didnt mention it clearly or rather not mention it at all.
I'm looking at on chip variations. So, modelling the entire design at SS or FF is too pessimistic. So, within the same chip, vt etc of the transistors are going to vary. I can take this into account if I do a statistical modelling (both at SS and FF). So, I get the delays as a probability distribution.
So, for a given value of vt, transconductance etc, I run spice and get the delay. Now, I vary these parameters accroding to some probability and get various delays with various probabilities. This way I can take care of on chip variations probabilistically.
Now, the issue is getting into the exact details of the modelling. Every parameter vt, cgdo, cgso, gm, idss, AS, AD etc will vary. The idea is to take uncorrelated parameters as much as possible. Also if I take too many parameters, its difficult to model. So I would like to take some 2 or 3 dominant guys. I'm thinking of vt and transconductance. Are these choices ok??
I hope you mean by using spice. Any other way is daft.
Standard i.c. design practise is to do worst case runs of strong/fast, nominal/normal and weak/slow. See for example,
formatting link
and
formatting link
You need to vary the mosfet capacitances as well.
If you use SS, you only only have to specify two special "multiplier models", and runs are all automated for that device. You don't have to work it out yourself.
multiplier are values that multiply nominal values, e.g.
It can be, but it it don't really matter if you can actually achieve a design that meets spec anyway. You usually only need to try and fine tune the calculation if the worst case design looks like the design wont make it.
I personally use that fact that wc can give a pessimistic answer to give me more confidence that the design will work for the conditions that are either not checked, or the models themselves are not accurate enough.
My personal view is that this approach (MC) is not that wonderfull. The reason being that, imo, one needs very large numbers of runs to account for all conditions. Transient runs can take a long time.
That's why one often just does worst case. Its a compromise between simulation time and accurate results.
I can tell you one thing though, in many companies, unless you do a wc set of runs, you wont get a sign off from your manager to get masks made.
What ever you do n addition, realistically, it *must* pass WC (not MC) to go to fab. End of story.
I noted in the other post what *are* the dominant ones to consider, you probaly missed this point being made, so I will repeat it:
These are the wc ratios (3 sigma) provided by a well known fab house.
ElectronDepot website is not affiliated with any of the manufacturers or service providers discussed here.
All logos and trade names are the property of their respective owners.