Hello Brecht, first, this is probably the wrong group for the topic you brought up. Yet, I'll try to share my limited experience with this (up to 18GHz). A transistor RF model is usually made of 1) the core base band model, and 2) RF "envelope" of parasitics around the active device, which is meant to model the BEOL interconnect. The baseband model is fitted first with BSIMPro+ or ICCAP. When you get a foundry PDK, the RF part of the model is often interpolated between the pre-solved geometries. Those geometries are better to have been characterized. But instead, they are usually field-solved using a "calibrated field solver." Here the word "calibrated" means that a few geometries were once characterized and the field solver parameters - such as the dielectric constant for example - have been fudged to give a matching solution. It's no surprise that at high frequencies you have difficulties even if you perform the de-embedding correctly. Before you begin to de-embed, you make an assumption about the equivalent circuit that represents the BEOL jig around your DUT, which connects the probe tips to the DUT. The fieldsolver (Momentum in this case), you use for emulation, "makes its own assumptions" by not taking into account some modes. There was a paper from IBM and Sonnet about resolving discrepancy between the measurements of a spiral inductor and the field-solver predictions. (Search Sonnet's website for the paper.) They found that the gap between the measurements and simulations is closed once the leakage resistance thru the p- substrate to the chuck is taken into account. Also, field solvers vary greatly in how well they perform numerical de-embedding. It is another significant source of error. I am not the most qualified person here on this list to explain this. Hopefully, the right people will see this thread and chime in on the subject. Also, at 100GHz I would think that a simple open-short de-embedding is not sufficient. You may have to assume several Z and Y sections alternating between the probe tip and the DUT. Hence, the de-embedding would involve several peeling layers. With sufficient number of elements / parameters you should be able to fit anything. At that point the accuracy of your measurements and how well you de-embedded to the probe tips become the source of error. NIST website should have good papers on this topic. Also, SUSS (in Dresden) has much experience in this and is worth talking to. -Vadim On Tue, Apr 13, 2010 at 3:50 AM, Brecht Machiels < brecht.machiels@xxxxxxxxxxxxxxxx> wrote: > Hi, > > I would like to obtain measurements of MOSFETs that are valid up to high > frequencies (110GHz). Unfortunately, I'm not familiar with de-embedding. > I'm currently investigating a number of de-embedding techniques: > > http://ieeexplore.ieee.org/search/srchabstract.jsp?tp=&arnumber=830987&queryText%3D.QT.A+four-step+method+for+de-embedding+gigahertz.QT.%26openedRefinements%3D*%26searchField%3DSearch+All > > http://ieeexplore.ieee.org/search/srchabstract.jsp?tp=&arnumber=4141067&queryText%3D.QT.Efficient+De-Embedding+Technique+for+110-GHz+Deep-Channel-MOSFET+.QT.%26openedRefinements%3D*%26searchField%3DSearch+All > > I've given the first of these a quick try by emulating a measurement > using a test fixture (unlikely to be optimal) and dummy structures > simulated in ADS momentum. Afterwards, I insert a MOSFET model into the > test fixture using HSPICE. From the obtained S-parameters I can perform > de-embedding. > > For large MOSFETs (W > 32 um), the de-embedded S21 and S12 are indeed > very close to the model's S21 and S12. S11 and S22 are less closely > matched to the model's, but still close. At higher frequencies (> 80 > GHz), the error becomes larger. > > For small MOSFETs, the de-embedded measurements only fit the model at > lower frequencies. I assume this is to be expected as the DUT's > properties become smaller compared to the test fixtures. > > Also, the articles on de-embedding I have read/scrolled through always > use large MOSFETs for the DUT (W > 64 um) when verifying the presented > de-embedding technique. > > This makes me wonder how typical transistor models are developed. I > assume a number of large transistors are measured to which a model is > fitted (BSIM, for example). Smaller transistors are modeled by means of > an implicit extrapolation. Could that be anywhere near how it's done? > > Unfortunately, I need to use small transistors (W as small as 2 um) in > my design, so I'm not sure how to continue from here. Any suggestions? > > Kind regards, > Brecht > > ------------------------------------------------------------------ > To unsubscribe from si-list: > si-list-request@xxxxxxxxxxxxx with 'unsubscribe' in the Subject field > > or to administer your membership from a web page, go to: > //www.freelists.org/webpage/si-list > > For help: > si-list-request@xxxxxxxxxxxxx with 'help' in the Subject field > > > List technical documents are available at: > http://www.si-list.net > > List archives are viewable at: > //www.freelists.org/archives/si-list > > Old (prior to June 6, 2001) list archives are viewable at: > http://www.qsl.net/wb6tpu > > > ------------------------------------------------------------------ To unsubscribe from si-list: si-list-request@xxxxxxxxxxxxx with 'unsubscribe' in the Subject field or to administer your membership from a web page, go to: //www.freelists.org/webpage/si-list For help: si-list-request@xxxxxxxxxxxxx with 'help' in the Subject field List technical documents are available at: http://www.si-list.net List archives are viewable at: //www.freelists.org/archives/si-list Old (prior to June 6, 2001) list archives are viewable at: http://www.qsl.net/wb6tpu