The distortion introduced by ADC's is intrinsically very different from that introduced by typical analog circuitry. ADC's have a transfer function which is highly non-linear at the small-scale (partly because of the quantisation steps intrinsic to the ADC process, but mainly because of the dynamic non-linearity and the structure in the integral non-linearity). On the other hand normal analog circuitry has a very smooth transfer function on the small scale and a smooth transition into saturation (an exception here would be cross-over distortion class AB or B amplifiers). Put another way, if we tried to describe the transfer function by a polynomial then an ADC would require a polynomial with an order up to about the number of distinct levels in the ADC (i.e. for a 16bit ADC we would need a polynomial of order 65000), whereas a good analog circuit used below saturation would only need a polynomial of order less than ten.
This fundamental difference is the reason why distortion products in an ADC are roughly constant in absolute power level (i.e. dB relative to full scale or dBFS) as the signal level is increased as you noted in your 2-tone tests. This can often be seen in the data sheets, for example p. 20 of the TI AFE8406 dual digital receiver, or p. 20 of the Analog Devices AD9446 ADC. So agreed, TOIP has little use for ADC's where the transfer function is dominated by higher order distortion. In contrast, for an analog circuit the absolute level of the major distortion products increases *faster* than the signal level as the level is increased, and hence the concept of TOIP does make sense for this type of circuit. One way of effectively "smoothing out" the transfer function and to reduce the level of the distortion products is to add a dither signal. The effect here is to randomise the errors that occur for a given input signal level. Without dither and for a certain input voltage the ADC output will be in error from the ideal infinite resolution case and this error will always be the same. However, if we add a changing noise-like signal then successive suamples of the ADC output will have different uncorrelated errors. The fact that they are uncorrelated implies that they will not appear in a spectrum as a distint spur anymore but as noise smeared out over the whole bandwidth, i.e. it will add a small amount to the noise floor (if the dither signal is itself a wideband noise signal then it to will raise the noise floor by an amount dependent on its relative level). I think the 16 MHz weak tone you added in your investigation is acting as a dither signal. In this case, since the dither is not a random noise signal, the power from the distortion products will not be spread evenly over the whole bandwidth but will be in a number of weak spurs. Still the effect will be to reduce the third order IMD level. There is a good article by Walt Kester on the Analog Devices website which explains why dither can improve the SFDR of an ADC: http://www.analog.com/library/analogdialogue/archives/40-02/adc_noise.html If the application we are interested in is SDR here for shortwave reception using an ADC followed by digital down converter (i.e. not the IQ mixer plus audio ADC approach), then my guess it that dither is completely unnecessary since the input signal will intrinsically contain a lot of noise and uncorrelated weak signals spread all over the input frequency range and these will effectively provide the desirable dither function. In fact, it even may be desirable from this point of view not to restrict the ADC input bandwidth by filtering too much to ensure there is enough noise to dither out the ADC small-scale nonlinearities! For the direct conversion approach the performance of a (24bit) audio ADC may be good enough to go without dither at all. Nick, G4JNX. Andy Talbot wrote: > I've placed the test description and results of some SDR-IQ linearity > tests (in .PDF format) on my website given below. Look in the SDR > projects section. > > www.scrbg.org/g4jnt/SDRProjects.htm > > If one thing comes out of this test - it shows that the concept of > using Third Order Intercept Point (TOIP) as a measure of performance > for wideband digital receivers would appear to be completely flawed. > > For example, adding in a third tone F3, at a lower level while > monitoring the 2.F1 - F2 Intermod product, causes that product to drop > in amplitude dramatically. And why does the IMP stay reasonably > constant (in level, not dBc) as the input level of the two tones is > changed. > > Explanations please ! > We're well into new territory here. > > Andy G4JNT > www.scrbg.org/g4jnt >
