The Designer's Guide Community Forum

The Designer's Guide Community Forum
https://designers-guide.org/forum/YaBB.pl
Simulators >> Circuit Simulators >> Spectre +turbo Simulations slower than Spectre Simulations
https://designers-guide.org/forum/YaBB.pl?num=1274077560

Message started by cmos.analogvala on May 16^th, 2010, 11:26pm

Title: Spectre +turbo Simulations slower than Spectre Simulations
Post by cmos.analogvala on May 16^th, 2010, 11:26pm

I have been trying to reduce simulation time of my netlist for quite some time.

I tried many things from moving to MMSIM to reducing netlist itself. We came to know that with Spectre +turbo simulations can run faster.

We tried to simulate a long netlist using spectre and spectre +turbo. on the same OS and the same machine. We found that Spectre +turbo was slower than spectre. From the log file it appears that spectre+turbo is taking very small steps, hence it's slower than spectre.

Any clues why so ? Is there a way to improve step size of spectre + turbo simulations? Is there any other way of getting faster results without compromising accuracy ?

Thanking You,
CA

Title: Re: Spectre +turbo Simulations slower than Spectre Simulations
Post by Marq Kole on May 26^th, 2010, 2:58pm

What Spectre +turbo does differently from regular Spectre is that the device evaluations are run multi-threaded. This encompasses a little overhead, so normally this is switched on when there are at least 256 devices per thread. Also the devices have to require evaluation - in generally this concerns mainly semiconductor models (BSIM, PSP, etc.).

If you do not want to compromise accuracy, there is generally very little you can do in Spectre +turbo. You could try to run with errpreset=liberal, but that reduces accuracy a little - it might be acceptable for your design/simulation.

Another alternative is to consider Virtuoso APS which is a true parallel simulator from Cadence, or look to high performance simulators from other vendors such as AFS from BDA or FineSim from Magma, to name just a few.

Cheers,
Marq

Title: Re: Spectre +turbo Simulations slower than Spectre Simulations
Post by Geoffrey_Coram on May 27^th, 2010, 5:54am

spectre +turbo shouldn't be slower -- Cadence R&D might be interested in your test case.

Have you tried other test cases? Can you break off blocks of this netlist and see if you can narrow down where the problem is?

Title: Re: Spectre +turbo Simulations slower than Spectre Simulations
Post by ywguo on May 27^th, 2010, 8:37am

Hi Marq,

This is interesting. I want to know what Spectre + Turbo do in the simulation. Does it run the device evaluation multi-threaded only? I think the device evaluation is done at the begining of a DC/AC/tran analysis. So it cannot reduce the simulation time significantly. Is it correct?

Yawei

Title: Re: Spectre +turbo Simulations slower than Spectre Simulations
Post by Geoffrey_Coram on May 28^th, 2010, 7:08am

ywguo wrote on May 27^th, 2010, 8:37am:

I think the device evaluation is done at the begining of a DC/AC/tran analysis. So it cannot reduce the simulation time significantly. Is it correct?

Yawei - you're confused. Device evaluation is done for every iteration of every dc sweep point or time point, so speeding up evaluation would certainly speed up dc/tran analysis.

(It might not be a big deal for an ac analysis, since the results for one dc op pt are used for all the frequency points of the sweep, and speeding up evaluation only helps with the dc op pt.)

Title: Re: Spectre +turbo Simulations slower than Spectre Simulations
Post by ywguo on May 31^st, 2010, 12:27am

Thank you, Geoffrey_Coram.