Well, you definitely get it right ! Your clock diagrams really look like those in the "M series manual", I really admire both your analysis and your ASCII art capacities
You were indeed right, I was using ticks as units for delay, and the convertclock is the 20 MHz onboard clock. I think I get it all now, thanks to your eplanations and my experimentations and as you mean it, I guess I could play with convertclock frequency and delay and find the "fitting" parameters set, I'm however afraid that this will be rather a mess to compute it for any number of channels at any sample rate (anybody ?) and I'm not willing to pay this additional complexity. I'll chose to read an uneeded AI first and ignore it. Since AO settling time is specified to 2 us and the max sampling frequency is of 250 KS/s with four channels, the second channel I read will be well timed with respect to AO. Of course, I'm loosing a few KS/s but I still sample fast enough for my application.
Maybe this situation should, however be dealt with by NI : when you use the same clock of AI and AO, obviously synchronizing them, the optimal convert delay could be computed including AO ? I've found this situation really didactic but rather annoying, I can't think I'm the only one who ever got it.
Anyways, thank you very much McDan for your time.