transient noise analysis

Ali,

The reason why there is a parameter is because a circuit simulator is not real life - you are having to attempt to represent continuous time with a finite number of discrete time steps. These time steps have varying spacing to try to get good accuracy at sufficient speed.

If you had noise at infinite frequency, you'd have to have infinitely short time steps. So by telling the simulator the maximum frequency, that sets the maximum timestep that the simulator can take - the higher the frequency, the shorter the timestep (it may take shorter timesteps than the maximum noise frequency dictates in order to follow the waveforms within your specified tolerances). Also, the transient noise assumes that the noise is white, based on the values at noisefmax, unless you also specify noisefmin - in which case it can then colour the noise (see "spectre -h tran" or the Spectre User Guide or Spectre Reference for more details - run "cdnshelp" to see these).

In general you'll get the best accuracy using a small signal analysis (such as noise or sp), or if it's a period circuit you could use pnoise/qpnoise/hbnoise or psp/qpsp. Transient noise is best for cases where either the circuit is non-periodic or has a large signal response to the noise - because it's not the fastest approach to simulating noise. If you have high noisefmax, you need short timesteps, and if you have low noisefmin, you need a long simulation to cover a period of the minimum frequency; both of these together can be a good way of getting a slow simulation - although APS can really help in this respect.

As for what you should set it to, you have to have some  understanding of the bandwidth of your circuit and any noise folding that is going on in the circuit. Hard to give any more specific guidelines than that...

Regards,

Andrew.

 Thank you very much for your time and the detailed description.

From what you mentioned the transient noise is therefore best for non-periodic and larg signal circuits, when noise is white. I think my circuits have the same situation.

Is the filcker noise considered in this analysis?

I have already optimized this design using noise ananlysis, and tried to get V**2/Hz < -138dB at all corners. Now when simulating using transient noise, I see that the output noise is highly depends on the noisefmax. For sure the circuit has finite bandwidth due to parasictics. It is a part of bigger design with discrete-time circuits with a sampling frequency of fs=5MHz (other parts of the design are verilogA models and do not generate noise). Changing noisefmax from fs to 128 fs procudes confusing results at different corners as following (noisefmin is not set).

Corners    tt      ss       sf        fs        ff              noisefmax
SNR    76.9    82.2    75.5    84.3    84.6    dB    fs = 5MHz
SNR    82.3    71.1    72.3    70.2    75.3    dB    16xfs
SNR    73.9    71.0    73.1    78.2    73.1    dB    128xfs
 

In all the simulations the maximum step size is 1ns, which is much less than 1/fs. However, in some corners increasing the noisefmax reduces the total output noise. Moreover, for those corners that output noise has increased vs noisefmax, trends are not similar.

I wonder if this bahavior can be explain for one circuit. It seems simulator is changing the paramters nonlinearly based on noise bandwidth and makes it difficult to trust the reseults.

Any idea?

Thanks,

Ali