The Designer's Guide Community
Forum
Welcome, Guest. Please Login or Register. Please follow the Forum guidelines.
Mar 28th, 2024, 11:08am
Pages: 1
Send Topic Print
question about MonteCarlo simulation strategy (Read 4671 times)
subgold
Community Member
***
Offline



Posts: 97

question about MonteCarlo simulation strategy
Aug 17th, 2010, 1:55am
 
let’s say I do a 1000-run MonteCarlo simulation to check the performance of a circuit, and find some malfunctions, which occur at only very few runs (maybe in 1 or 2). then I rerun the simulation at that specific run to figure out what the problem is, and find out a solution.

It could often happen that the solution is inadequate, or I didn't understand the problem completely, the error could disappear in that run number but appear at another one, after the netlist change. therefore, in order to check if the solution works, I need to do again 1000 runs of MonteCarlo simulation, or even repeatedly.

Furthermore, more often than not, I don't expect to solve the problem all in once, but have to do some debugging instead, i.e. to exclude all the potential errors one by one. In this case, I also have to run many times of 1000-run MonteCarlo.

in either case, the MonteCarlo verification becomes very time consuming and inefficient, owing to the iteration of the 1000 runs. so I am wondering if anybody has encountered the similar situation, and has some constructive suggestion to improve the efficiency of the MonteCarlo verification, either from a general design methodology view, or from a technical point of view (e.g. are there any tricks to maneuver in the simulator, etc.) and would appreciate any advise.

thanks in advance.
Back to top
 
 
View Profile   IP Logged
ACWWong
Community Fellow
*****
Offline



Posts: 539
Oxford, UK
Re: question about MonteCarlo simulation strategy
Reply #1 - Aug 17th, 2010, 9:22am
 
Latin hypercube sampling is available in some Monte-carlo engines, which in theory give better coverage with fewer runs.
Search this forum for details.

In anycase it important to understand what the Monte-carlo is doing in your technology i.e. It maybe that process parameters are not correlated when they should be so you could be over designing or vice versa. Look at the distributions you get, does it match the process data ?

Other things to try to get to grips with is the device failing; is it process skew related or process mismatch related? Mismatch in the kit may be pessimistic if you work hard on the phyiscal implementation, or even just change device sizing (for the same nominal performance) can improve your yield.

Often its a designers decision as to what is good enough and what level of simulation verification is good enough not to constitute a yield hazard... very few ICs/SoCs have 100% yield in production ... although i did work on one with >100% but that a different story involving some creative accounting  :)

cheers

aw
Back to top
 
 
View Profile   IP Logged
ywguo
Community Fellow
*****
Offline



Posts: 943
Shanghai, PRC
Re: question about MonteCarlo simulation strategy
Reply #2 - Aug 25th, 2010, 7:25pm
 
Hi subgold,

I don't think 100% yield is the goal when you run monte carlo simulation. Usually, I run monte carlo to get the distribution of a few parameters. If the 3σ of that parameter is within spec, I am pretty glad.

Yawei
Back to top
 
 
View Profile   IP Logged
Mooraka
Junior Member
**
Offline



Posts: 20
Lexington, KY
Re: question about MonteCarlo simulation strategy
Reply #3 - Aug 26th, 2010, 6:33am
 
In the simulation point of view, I think you can set simulation options to run only 2 monte-carlo sims at the edges of Gaussian distributions (set limit to 3sigma or 4sigma based on design requirement) of process params. If your ckt meets these 2 conditions then it meets all 1000 runs.
Back to top
 
 
View Profile   IP Logged
pancho_hideboo
Senior Fellow
******
Offline



Posts: 1424
Real Homeless
Re: question about MonteCarlo simulation strategy
Reply #4 - Aug 26th, 2010, 6:36am
 
Mooraka wrote on Aug 26th, 2010, 6:33am:
In the simulation point of view,
I think you can set simulation options to run only 2 monte-carlo sims at the edges of Gaussian distributions (set limit to 3sigma or 4sigma based on design requirement) of process params.
If your ckt meets these 2 conditions then it meets all 1000 runs.
Your approach is no more than "Corner Analysis".

For example. Agilent GoldenGate have following four algorithms as "Sampling Method Options for Monte Carlo Algorithm".      
(1) Regular(=Standard)      
(2) LHS (Latin Hypercube Sampling) ; Sampling that will indicate the characteristics of the distribution in 1/5th or less as many trials as are required when using ordinary Monte Carlo.      
(3) HHS (Hammersley Sequence Sampling) ; compared to LHS, even more efficient, requiring fewer samples.      
(4) Boundary ; similar to corner analysis, minimum and maximum values of the random variables are used.      

Back to top
 
 
View Profile WWW Top+Secret Top+Secret   IP Logged
subgold
Community Member
***
Offline



Posts: 97

Re: question about MonteCarlo simulation strategy
Reply #5 - Aug 30th, 2010, 4:43am
 
thx for all the above replies.

to be clearer, i am not targeting at 100% yield. for most of performance parameters, i also use MC sim just to obtain the distribution. 1000 runs without failure are more about the functionality checks such as startup, power on reset, threshold voltage monitoring, etc. for a product-oriented design, i need to guarantee the chip is always working even at some extreme conditions such as very low/high temperature, although some performance degradation is tolerable.

on the other hand, i think 1000-run is not overwhelming from quality control point of view. in fact the commonly-used 4-sigma standard makes complete sense only if you do at least 15,788 runs, although we neglect that fact most of the time and do some approximation from less runs, let alone some design which require 6sigma standard.
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.