Based on useful comments of The Photon, Andy aka and PlasmaHH.
As I got, there is indeed the parasitic low-pass. There is no magic bullet that will mitigate it. Large receiver capacitance is to kill the high frequency signal.
So the question, what is the maximum allowable load capacitance for CML (/LVDS) for say 1G, 5G and 10G Hz signals? If based on experience or supported by links - better.
How to design a high speed CML driver given a low-pass filter formed by parasitic load capacitance (several pF) and terminating resistor (50 Ohm)? The problem is that high frequency signals are attenuated by the low-pass.
1. I am designing a high speed (several GHz) CML driver.
2. The transmission line (a microstrip or a coax) must be terminated (either at source or at receiver) with 50 Ohm (assuming single-ended termination).
3. A parasitic capacitance of several pico-Farads is expected (either at the receiver or at the source) (contact pad's parasitic capacitance to ground).
4. The driver drives high frequency AC current. Terminating resistor and load capacitance form an RC low-pass filter.
For my specs (50 Ohm, 5 pF) signals above around 1 GHz are attenuated substantially by the parasitic filter. What can be done?
They say (URL, table 2), that CML may drive signals up to 10 GHz while i am stuck with around 1 GHz given 3 pF load. Do they assume very low (say 500 fF) load capacitance for such high frequencies (10 GHz).
Below is a simplified schematic of the driver in LTSpice (can be downloaded here) that shows how the filter works. Green line correspond to output voltage signal for 1 GHz and red one (attenuated) for 3 GHz input signals.
I can mitigate the problem at one end of the line by decreasing Rterm. However, decreasing Rterm at both ends is not an option (improper termination). So, the only option I see is to decrease substantially parasitic capacitance where the transmission line is terminated.
As Andy Aka kindly suggested I added 7.5 nH in series with the TL without noticeable improvement in amplitude or edges.
Below are test setup, simulation in Spectre with 7.5 nH and without it respectively.