Xref: utzoo comp.lsi:888 comp.lsi.cad:326 Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!cs.utexas.edu!oakhill!oakhill-gateway!rxet30 From: rxet30@toto.oakhill.uucp (Steve Hamm) Newsgroups: comp.lsi,comp.lsi.cad Subject: Re: Circuit Simulator Benchmarks Message-ID: Date: 22 Nov 89 21:08:18 GMT References: <5780@alvin.mcnc.org> Sender: news@oakhill.UUCP Distribution: usa Organization: Motorola SPS, Austin, TX Lines: 70 In-reply-to: kenkel@mcnc.org's message of 21 Nov 89 20:09:02 GMT In article <5780@alvin.mcnc.org> kenkel@mcnc.org (Stephen Kenkel) writes: > It has struck me recently that there does not seem to > exist a good, standard, publicly available collection of > benchmark circuits for circuit and timing level simulators. Nice idea. I hope there's some response. I put a request out for problem circuits some time ago and got, I believe, exactly six circuit files. (Mark Johnson, from MIPS, sent me four of the six circuits. Thanks Mark!) I also got a larger number of requests for my "library of problem circuits" -- but most of the circuits weren't really "problem" circuits, and six isn't a library. Anyway, I suspect there may be more interest in having such a library than in helping to build one. One place to start may be Berkeley. In the SPICE 3C1 distribution, Tom Quarles' writeup (actually an appendix to his dissertation) includes results from a number of "benchmark" circuits (with a fairly descriptive naming convention) that are said to be available from the Berkeley Industrial Liaison office. You might check to see if they wish to contribute these (and thereby reduce their distribution load). Anyone from UC Berkeley care to comment? As a footnote, if you're going to set something up, I'd personally prefer a mail server (ala NETLIB), since I don't have FTP available from my site. > At ICCAD, for instance, I noticed a number of talks giving the > performance of simulators, but on circuits > that had been selected by the authors of the programs. > This makes it difficult to qualify claims of "100X Spice" > and "always converges". Ah, the inevitable ICCAD shell game. I've found it interesting that in industry, at least, in spite of the many papers presented on simulators that all are 100X SPICE and "always converge" nothing seems to have displaced the direct method circuit simulators (e.g. SPICE). TANSTAAFL. Even with UCB standard SPICE input using 2G6 or 3C1 models, there will be games that can be played. Commercial simulators typically relax the SPICE 2G6 default convergence tolerances by factors of 20 to 500000. Some define convergence differently. And different versions of, for instance, the level 3 mosfet model (say, using nice smooth splines to remove nastyness in the I-V and Q-V characteristics :-)) aren't necessarily the "same" model as SPICE for all applications, but this massively changes the overall efficiency and convergence characteristics of a simulator. So an answer is not necessarily the "same" even when it (perhaps by fortunate coincidence) is close to correct. Still, having a collection of standard circuits would be far better than not. > DISCLAIMER: > MCNC distributes a circuit level simulator (CAzM), and > has a vested interest in the benchmarking of such programs. That beats a number of the commercial simulator folks, who may be more interested in making sure that such benchmarking doesn't occur ;-) Oh, yeah, my own disclaimer: I make my living hacking on a circuit simulator for in-house use. --Steve -- Steve Hamm Motorola Inc. Semiconductor Sector CAD, Austin TX (512) 928-6612 oakhill!monarch!rxet30@cs.utexas.edu