Hi,

I need some help explaining a "discrepancy" in my spectreRF simulation. I am working on reducing the thermal noise of the first stage integrator of a SD modulator. As is the norm for a newbie, I started out with an ideal model of the integrator and tried to do some experiments to gain some understanding. So here are the details of my setup. It is a very simple switched cap integrator, nothing fancy - the input caps are swapped in every phase.

-The implementation is fully differential. But the input voltage in the application is single ended. So the negative input of the modulator is always 0V.
-The amplifier is an ideal voltage-controlled-voltage-source of gain 1e7. And it is noiseless
-The switches have on-resistance of 1ohm and the input caps are about 350fF. These switches are noiseless again. Fs=10MHz
-There is no noise coming from any of the elements in the circuit. I add only a noise source in series with the input
-The noise source is made up of a noisy R and a C, the noise voltage across the C is mirrored using a voltage controlled voltage source of gain one. The output of this VCVS is applied in series with the input source. So the input noise source has power kT/C and this power is spread over the bandwidth defined by R and C. The bandwidth of the noiseless switches and caps are much higher than the noise source bandwdith so I'm not losing any noise power.
-At the output of the integrator there are two track-and-hold circuits cascaded and operating in a ping-pong fashion. The reason is that I want all my noise power at the output within the [0,Fs/2] band. I did not want to see broadband noise on my defined "output" nodes. So the final track and hold cap sees just sampled noise values.

My intention is to perform the pnoise analysis, input refer the noise power and get back the kT/C that I forced at the input. So here is what I do to get that.

1. Perform PSS, PAC and PNOISE (I use enough sidebands)
2. For PAC, I force an AC source at the input, and the transfer function of the whole thing is readily available.
3. I take the output noise density from PNOISE, divide it by the transfer function from PAC squared - I get the input referred noise density
4. I integrate the input noise density from 0 to Fs/2, and expect it to be equal to kT/C

So here is the thing I cannot explain - the resulting noise power I get at the input is in the vicinity of kT/C but depends on the bandwidth of the input noise source that I put in my schematic (the noise source bandwidth is always many times higher than Fs)
1. If I have a very large bandwidth noise source at the input in the schematic, the simulated input noise power comes out to be exactly 1/2*kT/C
2. As I reduce the bandwidth of the noise source, it gets closer to the expected kT/C
2. If I severely reduce the bandwidth of the noise source the input noise from simulation goes above kT/C

Why is this happening? Is it some sort of correlation effect between samples (smaller the bandwidth, more the correlation, and hence a smaller power source to generate the same power at the output)? I stick in a noise source of bandwidth defined by R and C, but when I refer it back using the above procedure, I always expect the noise power to be in [0,Fs/2] band. So are the correlation dynamics different and is that the reason?

I'd appreciate any insight on this.

Thanks and regards,
Sanjay

Hi Sanjay,

I think that after 4years you don't remember anything (I have a little memory ) but I want to ask you some info about your problem because is the same as mine.

Do you have fixed your problem?
If yes, how?

The ping-pong fashion, works as a ideal SAMPLER instead a SAMPLE_AND_ HOLD (the difference should be a sinc(x) shape of the output noise?

Many thanks,
Rino