Hi,

I'm designing a low frequency relaxation oscillator in Cadence spectre, using the new augumented pnoise algorithm and the shooting algorithm.  This oscillator is highly nonlinear with a really slow chargeup for roughly 6.6us duration and a extremely fast falling edge of about 1.1ns. The RC oscillator charges up from roughtly -5mV to 315mV on node "Va". I have been able to get pss to converge with the following parameters:

       'reltol  "3e-6"
       'vabstol "1e-8"
       'iabstol "1e-12"

as well as setting the parameters to

?lteratio "10" ?steadyratio "0.001"

in the pss options. These settings are well above the cadence defaults for the "conservative" setting.

I have noticed that the pnoise simulation gives different results depending on where in the period the pss simulation starts. I wrote an ocean script that picks a different Tstab every simulation run and I tell it to output the phase noise results at f0 / 20. These are the results:

(("104.5u"
       ("Va" 0.1467503 65449.94 "VmidP" 0.9907851
           -8190.086
       )
       ((7.110502e-11 8.277806e-15) 9.632774e-10 -101.481)
   )
   ("104u"
       ("Va" 0.1131296 68866.95 "VmidP" 0.9932706
           -2764.549
       )
       ((4.096261e-10 7.753837e-15) 2.312038e-09 -93.87612)
   )
   ("103.5u"
       ("Va" 0.07777253 72403.28 "VmidP" 0.993957
           -428.1294
       )
       ((1.607232e-08 5.677848e-15) 1.44824e-08 -77.93921)
   )
   ("103u" "Error")
   ("102.5u" "Error")
   ("102u"
       ("Va" 0.2978443 44353.19 "VmidP" 0.6551509
           -424485.7
       )
       ((1.793918e-11 7.262696e-15) 4.838403e-10 -107.462)
   )
   ("101u"
       ("Va" 0.2490793 52932.07 "VmidP" 0.9101286
           -131942.1
       )
       ((2.127916e-11 8.28333e-15) 5.269614e-10 -106.7205)
   )
   ("100.5u"
       ("Va" 0.2215643 56875.43 "VmidP" 0.9567804
           -64701.29
       )
       ((2.236783e-11 8.3611e-15) 5.402731e-10 -106.5038)
   )
   ("100u"
       ("Va" 0.1921638 60509.95 "VmidP" 0.9788811
           -29470.66
       )
       ((2.737521e-11 8.598666e-15) 5.976956e-10 -105.6264)
   )
   ("99.5u"
       ("Va" 0.1609975 63957.0 "VmidP" 0.9886011
           -12394.18
       )
       ((3.995595e-11 8.204394e-15) 7.220911e-10 -103.9842)
   )
   ("99u"
       ("Va" 0.1281264 67355.62 "VmidP" 0.9924951
           -4611.735
       )
       ((1.303551e-10 7.801459e-15) 1.304263e-09 -98.84872)
   )
   ("98.5u"
       ("Va" 0.09354027 70826.85 "VmidP" 0.9937917
           -1216.622
       )
       ((1.284531e-09 7.009538e-15) 4.094243e-09 -88.91255)
   )
   ("98u" "Error")

Where the last number (bolded) in each list is the phase noise in dBc/Hz.  the "error" entrys are where the PSS did not converge. "Va" is the slow chargeup, fast charge down node and "VmidP" is an internal comparator node. The two numbers following Va and VmidP are the voltage at t=0 and dV/dt at t=0. I specified "Va" as the oscillator node for the pss simulation. It seems like pss fails when tstab is chosen such that the pss guessing game starts close to the extremely fast falling edge of the oscillator,  and the phase noise result improves if we start later in the RC chargeup period. There is a nearly 30dB difference in these results for the same oscillator, even though PSS converged in each case. My question is, which result can I trust, becauase the results are so wildly varying?

Thanks in Advance