The lowest "noise" floor of FFT in Hspice is about -110dB due to my simulation results, and noise floor of output signal is only -65dB. You can refer to the pic below.

what does lead to the high "noise" floor in fft analysis, the circuit noise or simulator error? what about the value of the noise floor that we can accept for a design? and most important thing is how to reduce it to a acceptable level.

My amplifier works as signal conditioning before a 10bit ADC.

[quote author=vborich  link=1134893752/0#1 date=1135012119]Did you make sure that the output is in steady state? Try running the analysis for twice as long and see if the results change. [/quote]
Yes, I am absolutely sure the output is in steady state, and I can't remember how many times simulated.

And I have tried several methods to reduce the the "noise" floor, such as decrease Tstep in tran, tighten RELTOL option, choose more accurate transient algorithm. but the result does not change a lot although may be a little improvement due to some modifications.

Sometimes I think maybe there are problems with my ckt, for example, the circuit noise or something that result in poor fft noise floor. The output noise of my ckt is about 2uV/rt Hz, and it sounds resonable for a 4 stage high gain amplifier (maximum gain is about 60dB). what about your advice or do you encounter similar problem? How to resolve it or just can we neglect this phenomenon?

Thank you, and I have learned something from your reply. The fft "noise" floor can be decrease to about -75dB while increasing the tran run time from 10us to 50us. the reason???
The following is tran and fft statements:

**input nodes:in1 and in2
vid1 s1 m0 ac 1 0 sin(0 0.3m 35meg)   $$to keep output swing equal 1Vpp
vid2 m0 ic sin(0 0.3m 37meg)
evid1 s2 m1 m0 s1 1
evid2 m1 ic ic m0 1          $diff Signal use transformers
vic ic 0 2
rs1 s1 in1 50
rs2 s2 in2 50
.options acct post nomod brief dcstep=2 CONVERGE=1
.options fft_accurate acout=0 lvltim=3 reltol=1e-4  
.tran 0.2n 10u  $when use longer duration the lower "noise" floor of fft
.fft v(outn,outp) np=8192 window=kaiser alfa=3  $default start=0 and stop=10u in tran.
.end

althogh longer tran simulation time can imporve the "noise" floor, I don't know the limit due to this, and it takes long time (Typically 50us needs about 4-5hours).
Since the fft "noise" floor does not represent the actual noise of amplifier, I just need to lower the floor level enough to manifest the distortion. For example, IM3 for a given output is about -60dBc, then I just need "noise" floor -65dB or lower.
As you mentioned, If want to do tran with noise analysis simultaneously, what simulator can be used? I have SpectreRF and ADS, can one of them do this?

Just to be clear on my last post, you don't need to go all the way to 50 us. 10 us is sufficient if as you said you are confident that the circuit is in steady-state. In fact, if it appears that it's steady after, say, 5 us, then use 4us to 5us in your .fft statement.

Also note that SpectreRF can perform the PSS analysis using either shooting Newton (a time domain approach, good for circuits which require lots of harmonics) or harmonic balance, known as "flexible balance" in SpectreRF (a frequency domain approach, good for circuits which require a small number of harmonics - i.e. linear or weakly non-linear circuits). If you're using IC5141 USR3 together with MMSIM60 USR2, you'll see the choices appear on the PSS or QPSS choose analysis forms in ADE.

Regards,

Andrew.