The Designer's Guide Community
Forum
Welcome, Guest. Please Login or Register. Please follow the Forum guidelines.
Mar 28th, 2024, 1:18pm
Pages: 1
Send Topic Print
gmin value for functional verifications (Read 17287 times)
Dushyant Juneja
Junior Member
**
Offline



Posts: 19
Bangalore, India
gmin value for functional verifications
Feb 01st, 2013, 2:41am
 
Hi all,

I was just trying to make some endeavors to hasten up my full chip simulations, when I came to this question. I saw gmin set to pretty low values, its only 1 pS. Don't you think it might hasten up a bit better if I increase its value to something better, like 1 nS? I worry about the magnitude differences created in the solution matrix during the simulation, hence am trying for this.

Naturally, I am only bothered about a vague top level functional check. Accuracy is not too stringent here. And my chips are large.

Please put forward your comments. I don't know if I am significantly helping here.

* Dushyant Juneja
Back to top
 
 
View Profile   IP Logged
Ken Kundert
Global Moderator
*****
Offline



Posts: 2384
Silicon Valley
Re: gmin value for functional verifications
Reply #1 - Feb 1st, 2013, 9:35am
 
No.
Back to top
 
 
View Profile WWW   IP Logged
Dushyant Juneja
Junior Member
**
Offline



Posts: 19
Bangalore, India
Re: gmin value for functional verifications
Reply #2 - Feb 3rd, 2013, 10:00pm
 
Hi Ken,

Thanks for your reply.

So, am I correct in saying that simulation speed and convergence are not affected by how small or large the elements in Jacobian come out to be? I am asking because I see options for setting minimum and maximum resistances to be considered during simulation in certain recent fast spice simulators. Please clarify.

Thanks again for your support. Smiley
* Dushyant Juneja
Back to top
 
 
View Profile   IP Logged
Ken Kundert
Global Moderator
*****
Offline



Posts: 2384
Silicon Valley
Re: gmin value for functional verifications
Reply #3 - Feb 4th, 2013, 1:05am
 
;Gmin has little impact on speed.

I cannot comment on the timing simulator settings you mentioned (you referred to these as Fast-SPICE simulators, but generally "Fast-SPICE" simulators are not SPICE-class simulators at all and so the name is misleading). They are generally filled with undocumented heuristics and ao it is difficult to make general statements about them.

-Ken
Back to top
 
 
View Profile WWW   IP Logged
Dushyant Juneja
Junior Member
**
Offline



Posts: 19
Bangalore, India
Re: gmin value for functional verifications
Reply #4 - Feb 4th, 2013, 1:30am
 
Hi Ken,

Thanks again for the clarification. It seems then, that there is no "single" standard way to accelerate such simulators. Looks like I need to have a lot of iterations with the application engineer for fine tuning each of them.  :(

I really hope things go more generic in this regard.

BTW, I am checking on loosening the time step. Just wanted to check this with you. If I am interested in transients of period, say 1 ns, isn't it wise to set the minimum time step to something like 100 ps? This is opposed to default settings of 1 ps in general simulators, including Fast SPICE (I prefer to have an easier reference, though its a misnomer).
Back to top
 
 
View Profile   IP Logged
Dushyant Juneja
Junior Member
**
Offline



Posts: 19
Bangalore, India
Re: gmin value for functional verifications
Reply #5 - Feb 4th, 2013, 1:31am
 
Sorry, previous post was a typo... here's the correct one:

Hi Ken,

Thanks again for the clarification. It seems then, that there is no "single" standard way to accelerate such simulators. Looks like I need to have a lot of iterations with the application engineer for fine tuning each of them.  :(

I really hope things go more generic in this regard.

BTW, I am checking on loosening the time step. Just wanted to check this with you. If I am not interested in transients of period less than, say 1 ns, isn't it wise to set the minimum time step to something like 100 ps? This is opposed to default settings of 1 ps in general simulators, including Fast SPICE (I prefer to have an easier reference, though its a misnomer).
Back to top
 
 
View Profile   IP Logged
Ken Kundert
Global Moderator
*****
Offline



Posts: 2384
Silicon Valley
Re: gmin value for functional verifications
Reply #6 - Feb 4th, 2013, 9:29am
 
My recommendation is to start with the simulator's default tolerances (it is surprising how many people start off with the settings set very tightly by default). Then if you want things to run more quickly and your circuit is not overly sensitive, then set errpreset to liberal. If you need it faster than that, then see if there are things you can do to speed the simulation (turn off unnecessary high-frequency signals, avoid long time constants, etc.). If that is not enough, I start creating functional model for non-critical blocks.

-Ken
Back to top
 
 
View Profile WWW   IP Logged
Dushyant Juneja
Junior Member
**
Offline



Posts: 19
Bangalore, India
Re: gmin value for functional verifications
Reply #7 - Feb 4th, 2013, 9:40pm
 
Thanks Ken.

Alas, I wish things could be that simple. My constraints are that my team wants a full transistor level simulation only, since they often keep justifying against the behavioral models in their SPICE-Verilog cosims. This full chip run is like a side-test, for gaining a better confidence level and testing things like IR/EM analysis. Essentially, it seems to my understanding that I cannot go to the modeling path for such things. Please correct me if you think otherwise. Proceeding with the idea, I am hence seeking to get the maximum from my simulator. Hence am wondering on time step as an option, trying to look beyond simple preset simulation modes.

Secondly, can we change the topic of this post? I suppose we have significantly deviated from simply gmin discussions.

Dushyant.
Back to top
 
 
View Profile   IP Logged
Ken Kundert
Global Moderator
*****
Offline



Posts: 2384
Silicon Valley
Re: gmin value for functional verifications
Reply #8 - Feb 5th, 2013, 9:53am
 
The problem with timing simulators is that for circuits that contain analog they often get the wrong answer. If you are only interested in IR/EM analysis, they might be okay because in those situations you only need to approximate the behavior of the circuit to get roughly the right current levels, voltage levels, and operating frequencies. On the other hand, it can take a lot of time and effort to get them set up properly and become confident in their results. And in the end, they generally do not provide much speed-up for analog circuits.

A better option is the accelerated full circuit simulators, such as Spectre-APS and BDA's AFS. Presumably, you are already using these.

You are going to have a hard time doing better than default settings combined with errpreset=liberal unless there is something specific about your circuit that you can exploit, though what that might be I cannot imagine at the moment.

Really, only modeling offers any promise. And it offers a lot of promise when done well:
  • You develop the model and so control the approximations (unlike with timing simulations you understand the limitations of your simulations).
  • You can choose any speed up you want by the way you write your models (100x to 1000x speed ups are typical and 1,000,000x speed ups are possible in the right situations with the right models whereas timing simulation usually only provides 3-10x).
  • The simulations can start before the circuits are designed.
However modeling does require special skills and more time and effort than most people are willing or able to expend, so they often just do what they can with transistor level simulations, which is generally very limited, and then they move on.

There is a very clear trend in the industry towards the use of models for functional verification, with most all of the major companies heading that way, so I expect the reticence your colleagues are expressing will eventually diminish.

-Ken
Back to top
 
 
View Profile WWW   IP Logged
Dushyant Juneja
Junior Member
**
Offline



Posts: 19
Bangalore, India
Re: gmin value for functional verifications
Reply #9 - Feb 5th, 2013, 11:17pm
 
Ken Kundert wrote on Feb 5th, 2013, 9:53am:
The problem with timing simulators is that for circuits that contain analog they often get the wrong answer.

A better option is the accelerated full circuit simulators, such as Spectre-APS and BDA's AFS. Presumably, you are already using these.


Pardon me but I am not sure what you mean by timing simulators here. I am using fast spice simulators, specifically hovering between Ultrasim, Customsim (Hsim/XA) and Finesim. Do they differ significantly from Spectre APS? Request you to please clarify.

Ken Kundert wrote on Feb 5th, 2013, 9:53am:
Really, only modeling offers any promise.


Thanks for this. Of course, I am looking towards modeling techniques much more seriously now. That's one reason I am on your book a lot nowadays!  :)

Dushyant
Back to top
 
 
View Profile   IP Logged
Ken Kundert
Global Moderator
*****
Offline



Posts: 2384
Silicon Valley
Re: gmin value for functional verifications
Reply #10 - Feb 6th, 2013, 1:28am
 
The simulators you mention are not SPICE simulators, they are timing simulators. They call themselves Fast SPICE simulators as a marketing ploy. The primary distinction is partitioning. If transistor-level simulators partition the circuit into subcircuits and solve them individually, then they become flakey (unreliable and often inaccurate) and finicky (require a lot of user configuration to get reasonable results). Such simulators have been successful with particular types of circuits, such as memories, but in general do not work well for arbitrary analog or mixed-signal circuits. SpectreAPS does not partition the circuit. Nor, do I believe, does BDA AFS. Rather they parallelize the evaluation of the devices and the matrix.

-Ken
Back to top
 
 
View Profile WWW   IP Logged
Frank Wiedmann
Community Fellow
*****
Offline



Posts: 677
Munich, Germany
Re: gmin value for functional verifications
Reply #11 - Feb 6th, 2013, 2:19am
 
Dushyant Juneja wrote on Feb 5th, 2013, 11:17pm:
Thanks for this. Of course, I am looking towards modeling techniques much more seriously now. That's one reason I am on your book a lot nowadays!  :)

Another pretty good introduction to modeling techniques is chapter 3 "AMS Behavioral Modeling" of the book "Mixed-Signal Methodology Guide". This chapter is available in four pdf files at http://www.eetimes.com/electronics-blogs/eda-designline-blog/4397605/Book-excerp....
Back to top
 
 
View Profile WWW   IP Logged
Dushyant Juneja
Junior Member
**
Offline



Posts: 19
Bangalore, India
Re: gmin value for functional verifications
Reply #12 - Feb 6th, 2013, 3:21am
 
Ken Kundert wrote on Feb 6th, 2013, 1:28am:
The simulators you mention are not SPICE simulators, they are timing simulators. They call themselves Fast SPICE simulators as a marketing ploy. The primary distinction is partitioning. If transistor-level simulators partition the circuit into subcircuits and solve them individually, then they become flakey (unreliable and often inaccurate) and finicky (require a lot of user configuration to get reasonable results). Such simulators have been successful with particular types of circuits, such as memories, but in general do not work well for arbitrary analog or mixed-signal circuits. SpectreAPS does not partition the circuit. Nor, do I believe, does BDA AFS. Rather they parallelize the evaluation of the devices and the matrix.

-Ken


Thanks Ken. I was not expecting this with Ultrasim and alike. I suppose Ultrasim plays a part for the AMS designer analog kernel as well, isn't it? I think I saw this in one of the Cadence seminars at my company.

Next, could you kindly tell what technique does the Spectre APS core use, I mean, that "accelerates" it over conventional Spectre? Some kind of transistor approximations? I am not curious for the IP... I'm just wondering the algorithm... a bird's eye view, if you could present.

Thirdly, am I correct in understanding that you are recommending me Spectre APS for a large full chip IR analysis? It would probably take days to my understanding... (without behvioral models, ofcourse)

@ Frank: Thanks for the references Frank. I will study them in parallel.

-Dushyant.
Back to top
 
 
View Profile   IP Logged
Ken Kundert
Global Moderator
*****
Offline



Posts: 2384
Silicon Valley
Re: gmin value for functional verifications
Reply #13 - Feb 6th, 2013, 8:27am
 
It is my understanding that SpectreAPS is a conventional SPICE simulator with some relatively safe optimizations and support for multiple cores.

-Ken
Back to top
 
 
View Profile WWW   IP Logged
Dushyant Juneja
Junior Member
**
Offline



Posts: 19
Bangalore, India
Re: gmin value for functional verifications
Reply #14 - Feb 6th, 2013, 8:43pm
 
Thanks Ken, Frank for all your contributions. The discussion clarifies me significantly.

-Dushyant.
Back to top
 
 
View Profile   IP Logged
Pages: 1
Send Topic Print
Copyright 2002-2024 Designer’s Guide Consulting, Inc. Designer’s Guide® is a registered trademark of Designer’s Guide Consulting, Inc. All rights reserved. Send comments or questions to editor@designers-guide.org. Consider submitting a paper or model.