Hi ,
  I use Hspice's transient analysis to simulate a PWM circuit with LC low pass filter at output . Buy it always has "internal timestep too small" problem after all. I try to fix this problem by changing RMIN & TSTEP value , but it can not work . Anyone knows how to solve this problem ?

I use BSIM3V3 model form foundry and there is no behavioral model in my circuit . In my circuit , since it is a PWM circuit , so there are a couple of big NMOS/PMOS transistors at output side . I also put a LC low pass filter (L=33uH ,C=1uF) to filter out the PWM output  . Hspice diagnostic for nonconvergent are show as follows ;



                       node    subcircuit       old       new      error
                       name    definition    voltage   voltage   tolerance
                      (xt.sp)     t89           -7.996k    -6.996k  125.065
             (xt.xi228.pdrive) out_blk     5.363      5.343     3.748
                  (xt.r101:i1)  t89           -7.996k    -6.996k  125.065

    total voltage source power dissipation=  -18.5695        watts

          So what can I do for solving this problem ?

Added Item:

The inductor sounds like it is ideal.  Reduce the Q of that perfect inductor with some series resistance and parallel capacitance. Ideal reactive elements can do some strange things in simulators.

Jerry

I was suggesting you break the feedback loop just for diagnostic purposes, to make sure the system works correctly open-loop.

reltol = 0.01 is huge -- that means you're allowing the simulator to "lose track" of 1% of the largest current going into any/every node.

itl4=100 is also not always such a good idea.  You're telling the simulator to try 100 iterations before giving up and choosing a smaller timestep.  The default is 10, and most transient analyses average less than 5.  You could be making your simulation run a lot slower than it should.  It would be better to get to the bottom of the convergence problem.

Gear is more numerically stable than trap.  Ken's book (DG to Spice&Spectre) is one place to read up on the integration methods.  Or "google" the terms.

Luck?