mark@mips.COM (Mark G. Johnson) (04/09/89)
Sort of now and again one reads about improved algorithms for circuit
simulation, usually accompanied by results like " N times faster than
SPICE but just as accurate" ( usually the reported N is >3 ).
( In this note, I take "circuit simulation" to mean transient
analysis of _general_ networks of transistors, sources, and
passive elements ... maybe looking at the step response of an
opamp, maybe looking at the delay of a 32-bit adder in ECL,
maybe investigating the MSB-glitch when the inputs to a DAC
transition from 0111111 to 1000000, and maybe looking at the
behavior of a flipflop with near-coincident inputs (producing
metastability). In short, a _general_ simulator, not something
that handles *only* monotonic-waveform logic gates.)
Are any of youse (what a crummy word for 2nd person plural) out there
actively _using_ such a simulator? If so could you please post a
summary of your experience with such beasties? I'm especially interested
in the average, and the worst-case-ever-seen, runtime ratio between the
fast simulator and SPICE. What's the difference between average-
and worst-case behavior? Is the fast simulator faster for only
some kinds of circuits, etc?
(We're doing inductive-ground-bounce studies, with octal CMOS (FCTA)
chips driving an array of dram SIMM modules. We find that for one of
the flavors of SPICE, Berkeley 2G.6, the simulations run slowly.
Perhaps [only a guess] it's because every single node in the entire
simulation (both on- and off-chip) is inductively ringing at hundreds
of MHz so the "activity" is high and the timesteps are very short
[a guess].)
Thanks very much,
--
-- Mark Johnson
MIPS Computer Systems, 930 E. Arques, Sunnyvale, CA 94086
...!decwrl!mips!mark (408) 991-0208upton@ole.UUCP (Mike Upton) (04/10/89)
In reply to Mark Johnson's question about SPICE replacements: I am not familiar with any circuit simulators that can handle generic circuit transient analysis( OPAMP's, DRAM's, ECL etc) that are not based on SPICE-type sparse matirx optimization. There are a number of commercial products that are based on SPICE that have significantly higher performance that SPICE2G6. At Seattle Silicon we use HSPICE (from meta-software?). HSPICE is basically a Berkeley SPICE compatible simulator that has been re-writen to be faster and more robust. It has been so long since we have run 2G6 that we dont really have any valid performance comparison, but on non-trivial problems HSPICE is approx 3-10 times faster than 2G6. I think that even more important than the runtime is the fact that HSPICE is far more reliable and converges much more quickly than 2G6. We have no qualms about starting a 3-4 day spice run (when CPU cycles permit) and waiting for it to finish. We could never do that with 2G6. The improved convergance of HSPICE easily reduces the number of runs required to obtain an answer. Thus the overall throughput of the system is MUCH higher. I have a feeling that this is true with all commercialy supported versions of SPICE, I was just telling my experience. We run on apollos, SUN3-280's, and SUN4-260's, all of which give good performance. I would really like to run HSPICE on a MIPS-M2000 (is that what the R3000 based system is called?) Obligatory Postscript I have no ties to meta-soft (other than a user of their software) -- Michael Upton@Seattle Silicon (uucp: ...uw-beaver!tikal!ole!upton) /* Semi-conducting our business since 1983 */
mark@mips.COM (Mark G. Johnson) (04/12/89)
In article <560@ole.UUCP> upton@ole.UUCP (Mike Upton) writes: > >It has been so long since we have run 2G6 that we dont really have >any valid performance comparison, but on non-trivial problems >HSPICE is approx 3-10 times faster than 2G6. > And he's right. MIPS's production circuit simulator is HSPICE too, and we find it to be a bunch faster than 2G.6 as well. > >I would really like to run HSPICE on a MIPS-M2000 (is that what the >R3000 based system is called?) > Yep, M2000 has R3000 cpu. Runs 20X the VAX780 on HSPICE (see Performance Brief for details) ---- making life cozy for the circuit designers here. CMOS group has 3 M2000's and 12 M120's so there are adequate :-) cycles available for SPICE runs. Still, it'd be very interesting to hear about other, yet more extreme, simulators that go even faster. (If any). In particular, simulators that are faster on problems in which every node is jumping around like crazy (such as our octal-FCT CMOS driver ground-bounce studies). -- -- Mark Johnson MIPS Computer Systems, 930 E. Arques, Sunnyvale, CA 94086 ...!decwrl!mips!mark (408) 991-0208