IP3 is power level at which the EXTRAPOLATED Input vs Output Power Curve
CROSSES the EXTRAPOLATED Input vs 3rd Order IMD Power Curve, see Fig 3:
http://www.arl.army.mil/arlreports/2007/ARL-TR-4235.pdf
So, as I've discussed here and on my webpage, we try to MATCH the Preamp
to the application, so as to MAXIMIZE Spurious Free Dynamic Range (SFDR).
Unfortunately in many REAL systems, the 3rd Order IMD Power Curve doesn't
follow a well defined straight line and might exhibit a hiccup half-way thru,
making it difficult to determine where the EXTRAPOLATION line should lie.
Some amps also exhibit a hiccup in the Linear Input/Output Curve, with two
vice only one gain compression point. So the IP3 "Spec" point can be a very
sloppy exercise by the test engineer, trying to draw two straight lines across
what may actually be to very wiggly curves....which probably vary quite a bit
due to typical part-to-part variations.
And, BTW, the various IP3, IM3 and other spec points can vary with frequency
and input signal levels:
http://www.ti.com/lit/an/sboa077/sboa077.pdf
[There is usually a reason Preamp manufacturers are stingy with some specs.]
So I wouldn't want to use IP3 (a fictitious, non-realizable, power level well
above the max output power level) to guesstimate the actual levels of the
3rd Order IMD noise products (which are desired to be close to the very low
Thermal Noise floor).
The inherent non-linearity of the 3rd Order IMD Curve can be clearly seen
in Fig 3.14 and 3.18 for a
modeled AlGaN/GaN HEMT device. Also
note that 3rd IMD curve is least linear when biased for the "sweet-spot":
http://www.uni-kassel.de/upress/online/frei/978-3-89958-654-1.volltext.frei.pdf
Modeled MOSFET device (Fig 3) shows a highly non-linear 3rd IMD curve:
http://ieeexplore.ieee.org/xpl/articleDetails.jsp?reload=true&arnumber=1309692 [Requires IEEE access]
also measurements for a very low-noise Cascode MOSFET LNA (Pg 37):
http://rfic.eecs.berkeley.edu/~niknejad/ee242/pdf/ee242_lect13_mosamp.pdf