Quote:

Actually this problem was identified and resolved by the initial SpectreRF developers (Ricardo Telichevesky, Jacob White, and myself) before SpectreRF was first released.  We did not think much of it at the time and the idea seemed to have been lost after we all moved on to other things. It was Geoffrey that resurrected the idea while trying to resolve anomalies he was seeing in the flicker noise of various models and simulators.  He then recruited Colin and Kiran to work through all the details, verify the results, and write the paper.

-Ken

Geoffrey_Coram wrote on Jan 20^th, 2020, 5:12am:

Thank you for the confirmation. As all the major simulator vendors for IC design seem to be participating in CMC (http://www.si2.org/standard-models/), these models will hopefully be fixed very soon.

Geoffrey_Coram wrote on Jan 16^th, 2020, 6:36am:

Thank you for sending me your BSIM3 testcase. I can exactly reproduce your plot with the most recent Spectre version 19.1.0.237.isr3.

Geoffrey_Coram wrote on Jan 20^th, 2020, 5:12am:

You are probably right that it's not entirely fair. I had not been thinking of the fact that the BSIM4 code is only targeted at the original SPICE from Berkeley, which does not have any of the analyses that are affected by this problem.

I'm not sure it's entirely fair to say the error is in the original C code from Berkeley, since Spice doesn't support Pnoise and thus the model has no way of communicating the sign of the modulation.

The R3 code on nanoHub does have this error, as do the R2_CMC and R3_CMC models from the Compact Model Coalition (http://www.si2.org/cmc). I have update those latter two models and the beta code is being tested by other CMC member companies.

Dear subtr and Horror,

Coupling between on-chip inductors can be a very significant problem as I have had multiple experiences concerning on-chip inductors used in LC based VCO. There are multiple papers addressing the issue. Proximity is an issue, but the magnetic fields fall off rapidly and hence increasing spacing between on-chip inductors is of limited use if you are experiencing significant coupling. However, the source of the coupling is also related to the non-linear nature of the varactors of the VCO and coupling that is introduced to them. Hence, the effect is not one of only inductor to inductor coupling. It can also be related to numerous coupling sources into LC based VCO varactors. In my experience, extremely careful layout and experience with this type of interaction is necessary to avoid significant coupling.

Shawn

I downloaded the archive from the link given in the "What's New" section of https://designers-guide.org/ and ran the simulations with the most recent Spectre release. The result for BSIM4 is exactly the same as in your original post, so the problem still exists there. This is not very surprising because according to the paper, the error is already in the original C code from Berkeley. Nobody seems to have been aware of this issue up to now; if I'm not mistaken, the error also seems to be present in the R3 resistor model by co-author Colin McAndrew from 2014 (http://dx.doi.org/10.4231/D3QB9V64G).

SpectreRF is often creating little puzzles like this. It always takes some deep thought to come up with possible explanations and then some experimentation to verify them.

My theory is that width of the sin(x)/x curve in the noise results is set by the pulse width (1.5μs -> 666kHz), whereas the in the PAC results of figure 6 the output is resampled with a 2.5μs period and so has width of the  sin(x)/x is 400kHz.  I am not sure why the unsampled PAC output also seems to have a 400kHz width.

Unfortunately, I do not have the time to explore this question to the depth it would require to completely explain the results.

-Ken

I think it's odd that in BSIM3, the "old" noimod=1 (just using kf and af) gives the correct results. The new unified flicker noise model[1] (using noia, noib, noic) is noimod=2 (or 3) in BSIM3, and has incorrect results.

This is flipped from BSIM4, where the new unified flicker noise model, fnoimod=1, gives the correct spectrum.

[1] K.K. Hung, P.K. Ko, and C. Hu, "A unified model for the flicker noise in metal-oxide-semiconductor field-effect transistors." IEEE Transactions on Electron Devices 37.3 (1990): 654-665.

I can get a similar picture for BSIM3 from the simulator I use. I had to increase the gate voltage because the default threshold voltage in BSIM3 is a little higher, and we want to be sure that the device is in the linear region - you can actually see a bit of a bump at 2 f0 in the noimod=1 results, which I believe is due to the device going slightly out of the linear region.