Does anyone have benchmark results for spectre on 32-bit rhel ws4 compared to its 64-bit cousin,
on identical hardware of course.

I have a couple systems running various flavors of ws4 as shown below, but none on identical
hardware. I am wondering if it might be faster for typical spectre sims (few hundred to few thousand equations)
if the machines were running 32b ws4 instead of 64b.

fi0x:~> uname -a
Linux fi0x 2.6.9-5.ELsmp #1 SMP Wed Jan 5 19:30:39 EST 2005 i686 i686 i386 GNU/Linux

fi1x:~> uname -a
Linux fi1x 2.6.9-5.ELsmp #1 SMP Wed Jan 5 19:29:47 EST 2005 x86_64 x86_64 x86_64 GNU/Linux

Thanks for the info.

As far as I can tell, spectre (mmsim61), ultrasim, and assura 316 all have 64bit versions of the tools,
and these are the ones I care about the most.

Typical sims I might run with spectre would be something like a simple block
or behavioral model which I do a transient 2ms run on. It takes about 2 min
on a P4 3.8GHz machine. The sim is scripted to run between 50 and 100 cases.
For something like this, it's not a terribly big deal if 32b-spectre is slower than 64-b spectre.

Circuit inventory:
             nodes 116
         equations 196
    ahdl simulator 1    
    va_ideal_clamp 3    
   va_ideal_clamp2 3    
       va_sr_latch 3    
     va_comparator 3    
         capacitor 15    
              cccs 6    
             diode 15    
          inductor 3    
              mos1 12    
          quantity 9    
             relay 9    
          resistor 27    
              vccs 6    
              vcvs 6    
           vsource 47    

However, for mid-level or top-level simulations and extracted sims, using
spectreVerilog,  the netlist looks like:
Circuit inventory:
             nodes 4318
         equations 13684
               a2d 36
           bsim3v3 3999
         capacitor 2355
               d2a 53
             diode 928
          inductor 2
           isource 2
          quantity 9
          resistor 1911
              vbic 28
           vsource 650

Something like this takes between 10hrs and >40hrs for a 1ms transient run on
the same machine. The actual time depends on various modes of the circuit and
its dynamic behavior during certain intervals of the simulation.

I am looking for any possible way to shorten these 40hr runs to something more
reasonable. The machine is not in danger of running out of memory for these sims,
so if 32b spectre could be any faster, I'd retool my fast box for 32b linux instead.