The Designer's Guide Community Forum

The Designer's Guide Community Forum
https://designers-guide.org/forum/YaBB.pl
Simulators >> Circuit Simulators >> more about Xoomsys
https://designers-guide.org/forum/YaBB.pl?num=1116781857

Message started by 101questions on May 22^nd, 2005, 10:10am

Title: more about Xoomsys
Post by 101questions on May 22^nd, 2005, 10:10am

I posted a question about Xoomsys accelerator a few months ago.
Now they claim they have a working alpha version:

http://www.eet.com/news/design/showArticle.jhtml;jsessionid=WBOBXOCVOFLKEQSNDBCSKH0CJUMEKJVN?articleID=163106108

Anybody tried this one? Intends to try?

Title: Re: more about Xoomsys
Post by Baolin Yang on May 23^rd, 2005, 5:29pm

Quite some experts in parallel computation and fast circuit simulation have worked on parallel simulation in the 80s and early 90s.

Xoomsys' claim is that some method from control system can solve the hard problem. I am just not sure whether it is applicable to all types of circuits or just a few types of circuits. For strongly coupled complex analog or mixed signal circuit, it is hard to imagine that one can partition it in such a way that not much communication is needed.

Title: Re: more about Xoomsys
Post by 101questions on May 24^th, 2005, 8:49am

Yes, many failed to imrove spice simulator's speed without large sacrificions.
This time it's different. Xoomsys claim is that they are speeding up the existing commercial simulators like spectre, rather than writing their own one.
Assuming one spectre license costs like 30-40 Linux stations, one pays for single license and 30 station network and get power of 30 spectre simulators for the price of two licenses. Looks like a sweet deal.

Title: Re: more about Xoomsys
Post by Paul on May 25^th, 2005, 4:24am

I don't know their product, but I guess you didn't get the way of operation of license servers with floating licenses...

Paul

Title: Re: more about Xoomsys
Post by 101questions on May 25^th, 2005, 10:59am

Well, it's not like a license server, but if it speeds up a single-license simulation by a factor of 30, the simulation throughput becomes like for 30 separate simulators, right?

Title: Re: more about Xoomsys
Post by Baolin Yang on May 25^th, 2005, 1:34pm

All software vendors charge one license per computer. Oracle even insists on charging one license per core in the case of dual-core computers. Never heard of anyone who is willing to charge one license for users to run their software on 30 computers.

Title: Re: more about Xoomsys
Post by 101questions on May 25^th, 2005, 9:42pm

OK, let's put it another way. If Xoomsys technology works, it allows to use N licenses either for N simultanious simulations or for 1 simulation that runs N times faster. It's a very imprortant trade-off for me.

Title: Re: more about Xoomsys
Post by rf-design on May 31^st, 2005, 6:15am

Indeed,

in the past 15 years there where approaches to speed up spices modified nodal circuit simulation. The parallel approaches are all depending on the algorithms. The device evaluation for instance could be run in parallel but for big size circuits the sparse matrix is dominating. Spectre does have application optimized code for doing that job. Parallel sparse matrix and multirate integration with latencies is the key. I think there is simply a manpower limitation to improve spice from the algorithm side. The parallel sparse matrix code today also show that the vector machines does not run them sucessfull. They are more powerful for FEM problems. Message passing architectures have latency issues. There is some drive from molecular dynamics where the problems have a similar sparse structure. I guess that this market would benefit more for the circuit simulation in future.

From the practical side use a single license on a liquid nitrogen cooled 3.2GHz FX-55.

For multicase verification use the licenses of the company and spread them over a number of boxes.

Organize your problem size and the verification hierachy scope that you get your result overnight.

Title: Re: more about Xoomsys
Post by 101questions on May 31^st, 2005, 9:12am

I'm sure that if speeding up the spice required only applying some well known methods, the EDA companies would spend this effort many years ago. I agree with you that all the known techniques are already milked down to exhaustion.

My hopes are that Xoomsys uses something new and significantly different from others. Even their basic claim that they do not write their own simulator, but rather they are taking an existing binary code and run in in parallel on networked servers - this looks very different from other attempts.
I eager to learn more how they do it, or at least hear if it works from somebody who tried it.

Title: Re: more about Xoomsys
Post by rf-design on May 31^st, 2005, 11:47pm

I tried do catch up some informations about the technology but got nothing. I guess that they take a netlist and analyse how to separate the circuit. If the signal flow, analysed from the netlist, could be figured out as an directed tree then they could break up the netlist and simulate separate pieces. I am not shure that all supported simulators accept PWL sources for the interconnection signal flow with very high number of points.

This approach is similar to sparse matrix ordering methods but the difference is that from the netlist information you only know what is I, O or IO. And this categorisation is seldom true for analog.

So wait 14 days. Then they should leak some information.

Title: Re: more about Xoomsys
Post by rf-design on Jun 1^st, 2005, 6:50am

Here a short summary what Al Davis knowns about his own simulator performance limitations:

http://lists.gnu.org/archive/html/gnucap-devel/2003-10/msg00000.html

Title: Re: more about Xoomsys
Post by 101questions on Jun 1^st, 2005, 12:23pm

rf-design, thanks for the interesting reading.
I believe all the major simulator implement most of the tricks that Al mentioned and then some more. However, they all are stalled in terms of speed to where they've been some 10 years ago. It looks like Spice-like simulator developers gave up on speed problem with all the speed progress moved over to ultrasim and the likes.
Let's look if Xoomsys can change anything here.

Title: Re: more about Xoomsys
Post by rf-design on Jun 2^nd, 2005, 1:27am

I think they did not give up but the talent went a different way as the market does not deliver enough to justify deep research. There could be help from other application like chemistry and molecular dynamics. That first deliver efficient sparse matrix code for the upcoming multicore/multiprocessor machines which code be reused for circuit simulation.

I agree that on single CPU/Core there is not much left which justify new spice kernels.