THD is much harder to measure, never mind at the performance level that a sound card or a DAC is capable of. But there is a possibility that your mismatch was caused by something else. RMAA will give measurements that are, just like you say, 2-2.5dB different from each other, by simply setting your level differently when "calibrating" at the beginning. It will accept down to -3.5dB or so but it asks of you to set it to -1dB. If you set it to -3dB instead, which I like doing to reduce distortion, then you will notice your results are better by about the amount of the difference between your level and -1dB. So they don't compensate for that.
I was not referring to THD measurements, but rather noise and dynamic range. It can be tested easily by using the "Generate WAV" function of RMAA in 44.1 kHz/24-bit format, then adding noise to the WAV file (for example by reducing its effective resolution to 16 bits with TPDF dither using
this utility), and analyzing the modified WAV with RMAA. Other measurements, such as crosstalk, can be tested with the same generate->process->analyze method.
In the case of the dithered test signal, RMAA 6.4.1 prints these values:
Noise RMS: -96.7 -96.8 dB
Noise RMS (A): -97.8 -97.8 dBA
Peak value: -90.3 -90.3 dBFS
The peak level is correct, but the noise level should be -93.3 dB unweighted and -95.7 dB A-weighted, or -93.7 and -95.8 with a 20 kHz bandwidth. These are referenced to a 0 dBFS sine wave, which is what one would expect RMAA to use as the reference as well, especially since 0 dB on the graphs corresponds to 0 dBFS in the WAV file. In any case, a lower reference level would make the measured noise levels worse, not better. I think what RMAA really prints as "Noise RMS" is relative to a 0 dBFS square wave, that is why it is lower by 3 dB. That is not necessarily a problem, but this fact could have been made clear by the program. On the other hand, "Noise RMS (A)" was probably meant to be referenced to a full scale sine wave, but the
A-weighting filter in RMAA is not correctly normalized to have 0 dB gain at 1 kHz; that explains the 2 dB difference.