Each original sample (which are all retained in the Yggy's mega combo burrito filter) could be from 16 (Redbook) to 24 (Hi-res) bits. The 20 bit output samples are rounded off (not truncated) right before the DAC.
On a different occasion you've mentioned this, too:
http://www.head-fi.org/t/782824/schiit-fire-and-save-matches-bifrost-multibit-is-here/480#post_11988106
In Yggy, Gumby, and Bimby all trailing bits are rounded at 20, 18, and 16 bits respectively - not truncated which introduces errors.
But I still don't get it... Technically what is difference between truncation and rounding in this context? Why does truncation introduce errors, and how? I'm hoping someone could come up with a brief example... And how does this all fit in with Mike's/Jason's assertions that there is barely any actual content above 20 bits (2^20=1048576), even in 24 bit files...
truncation drops everything right of the "decimal point" (but in binary) in the smooth analog input conversion to discrete ADC steps
so -0.999... and +0.999... get truncated to 0 giving a "deadband" or "crossover distortion" 2 bits wide arround 0, while rounding makes all the steps just 1 bit wide everywhere
the truncation deadband generates more quantization noise, of different and annoying character than rounding - but both give "correlated quantization noise" - "spitty" distortion as the analog signal fades below the thresholds of the lower bits of the ADC in recording, DAC in playback
you can
hear it in the 8 bit audio samples at:
http://audio.rightmark.org/lukin/dither/
the next step is adding dither - as the Lukin page shows, gives listening examples - dither is near universally recognized as an improvement over rounding alone for digital audio, nearly all commercial CD will have been dithered in the last bit, most with "noise shaping" - for which Lukin is showing various options
dither gives a audibly smooth fade of the audio/music signal to below the lowest bit
Mike has gone ballistic before when it was pointed out that the practice of dither in commercial digital audio releases makes "bit perfect, dammit" a rather odd sounding affectation to those of us familiar with the math, the noise of microphones, playing venues, listening rooms... and the commercial digital audio production practice and Psychoacoustic hearing details
as someone experienced in designing Scientific and Industrial Instrumentation; selecting, designing, wiring ADC, and DAC to computers since the 1980's, I would dither inside the DAC when using R2R DAC chips of <24 bit resolution
with "flat/white" spectrum Triangular Probability Density Function (tpdf) dither the "cost" is ~ 3dB of added noise over the full bandwidth, with upsampling DAC less of dither noise is in the audio band even without noise shaping
the most trivial "1st difference" colored tpdf dither wouldn't move the Audio S/N spec meaningfully at 8x upsampling
but it would be "below lsb linear" - which the rest of the digital audio world has decided is the more important feature to have