Electronic – External Clock for AD7193 ADC

You might need to slow down the ADC. The analog to digital converter runs at half the system clock frequency by default so at 6MHz the ADC's input clock is at 3MHz. Setting ADCSRA |= 2 (for the ATmega328, datasheet section 23.9.2, page 265) divides the clock by 4 instead of by 2. 12MHz / 4 = 6MHz / 2 = 3MHz.

First thing you need to recognize is that designing with a 16-bit ADC is not trivial. Even at 1 sample/s, you need to pay extreme attention to every aspect of the design to achieve 16-bit precision, or even more difficultly, 16-bit accuracy. At 130 MSa/s, everything is even more difficult.

The parts you need to do this kind of design simply won't be inexpensive. First, because of the extreme precision and careful testing needed to achieve the required performance. Second, because this kind of thing isn't done in mass-market products, so the parts aren't built in the kind of extremely high volumes that can bring the price down for everyone.

As Dave says in another answer, be sure you really need 16 bits before you go down this road. But maybe you really need 12-bit precision, and you know that if you use even a 14-bit ADC you're going to have a hard time achieving that, so you're designing with 16-bit ADC and optimize everything else as much as you can.

Another key is likely to be understanding exactly what specs you need to make your system work, and don't over-specify your clock jitter. In an SDR application, you're going to be doing math on the samples to extract specific frequency bands, etc, which will have an averaging effect over many cycles. So you might not care too much that absolutely every sample is timed perfectly, only that over your calculation interval, there isn't too much deviation from ideal timing. How much is too much, of course, depends on what kind of math you're doing and how small a signal you need to extract from how much noise.

CTS Valpey, for example, has XO's with rms jitter specs as low as 200 fs. But this spec is defined when the phase noise is integrated over a specific frequency band, 12 kHz to 20 MHz (relative to the carrier). If the total cycle-to-cycle jitter is considered, the spec jumps to 3-6 ps, depending on the center frequency.

Let me also address one comment you made in your question:

OCXO are extremely stable over time ( years ) and are usually used for that.

The "ovenized" part of that product mainly reduces the drift due to temperature change in the surrounding environment, which can be significant over time scales of minutes or seconds, not just years. It will also reduce wear on the part due to thermal cycling and improve the stability on a time scale of years.

For the < 100 fs jitter range you're looking for, you might actually need an OCXO to prevent small temperature changes affecting the performance during the time it takes to measure the jitter accurately enough to know you've achieved your spec.