(A couple of users suggested I repost in this thread)
I bought a Rigol DS2072 to replace my 35 year old Phillips PM3212 that just died. I've been evaluating the High Res mode and it does work. I programmed my computers 16 bit DAC to produce a sine wave with a tiny 0.7 LSB blip on it, much like the Agilent video posted earlier (in the other thread). The Rigol could easily resolve the blip in High Res mode despite the fact that it was only 0.7 LSB of the 8-bit ADC.* The first figure shows the High Res trace blown up. This trace was taken 100 mV/div and 2 ms/div and blown up post acquisition to 5 mV/div. The sub LSB blip is easily resolvable. Of course, for a repetitive signal you could also just turn up the sensitivity to 5 mV/div and hunt for the blip with the offset and time/div and this is shown in the second figure.
Of course, the extra resolution you get from boxcar averaging doesn't mean much if the DNL (differential non-linearity) of the ADC is not good, so I used the ramp and histogram method to measure DNL. The results are that it's pretty good: only +-0.15 LSB (see third image). I also did a best line fit deviation and the error is somewhat larger at +-0.5 LSB. (This is very similar to INL, integrated non-linearity, but it wasn't exactly calculated by integrating the DNL, it's just the deviation from a straight line fit to the ramp signal. INL specs are always worse than DNL, but the refer more to absolute accuracy not the ability to resolve small signals.) Note these are upper limits as it depends on the test signal generated by my computers 16-bit DAC being perfect.
I also measured the noise on the 500uV/div scale as 80uV RMS in Normal acquisition mode and 40uV RMS with the High Res mode, both with the 20MHz BW limit on. With the 20MHz BW limit off and Normal mode I got 100uV RMS. These are all with the input shorted with a 50 ohm terminator.
*When calculating LSB remember that the 8 vertical divisions are not the full range of the ADC; it's actually 10.24 divisions with headroom off the top and bottom of the screen plot region.
Other notes/problems:
Inputting a pulse that should have a rise time of 5 ns (according to the specs for the chip I used) gives a rise time of 7.1 ns on the scope. That means the scope rise time is 5 ns which matches the 70 MHz spec.
The "Measure" box does appear to only work with the screen trace data, but the 6-digit frequency "Counter" works directly from the signal (with the trip point defined by the trigger level). On a 1 MHz input signal the Rigol "Counter" matched my external frequency counter by 10 Hz. (Of course the frequency counter is 40 years old so who knows if it's accurate.) On my scope the "Measure" box can get hung requiring a reboot. If I input a 1 MHz signal and resolve the cycles the "Measure" box gives correct data. If I slow down the time/div, say to see modulation of the RF carrier, the "Measure" box will get confused and quotes, say, freq > something. If I speed up the time/div to see the 1 MHz cycles again the "Measure" box will remain hung with bad data and nothing I tried gets it working again except a power down.
I was unable to get the scope to save screen captures to the USB stick from the "Storage" menu; it just saves a blank screen with no trace. It will save correctly to the USB if I press the "Print" button (no printer attached). It also saves data in CSV mode properly. Note that even if you take data in High Res mode the data saved externally in CSV mode is still 8-bit quantized. High Res cleans up the data saved, but it's still 8-bit and no extra precision shows up in the file (while it does show up on the screen).
All in all I like this scope a lot. Note that I'm an old person who was used to analog scopes and I've had a dislike for digital scopes for quite awhile, so I'm a bit hard on them.
Edit: added caveat