Schiit Yggdrasil -- 21 bit?


Schiit says that Yggdrasil is a 21 bit DAC. But the DAC chips that they put in the device ( Analog Devices AD5791BRUZ, 2 per channel) are 20 bit with the error of plus-minus 0.5 LSB.

How can the DAC be 21 bit if the chips are 20 bit? Using two chips per channel does reduce the RMS voltage of the noise by  a square root of 2. But how can you get to 21 bit from there?

Can someone please explain.
defiantboomerang
JA is super duper diplomatic or euphemistic as always.

Here is a five minute youtube that explains why dither is so important in digital audio.

https://m.youtube.com/watch?v=zWpWIQw7HWU
JA is wrong, Schiit uses rounding, not truncation.
JA believes that MQA (lossy, proprietary, licensed platform)
is " the birth of a new world".
The "extra" 1 bit in Schiit balanced multibit Dacs is deducted from the 2 phases +6dB.
If you watch the video you will see that rounding is practically the same as truncation and does similar damage.

Sorry but Schiit don’t understand what they are doing. This is common when good designers with a strong background in analog design start making digital components.
After a failed appeal to authority now  comes a referral to youtube and arguing ad hominem. Scores of other logical fallacies still at hand.
Rounding and truncation are similar, but with rounding being a broader function over truncation in this use. The idea behind rounding in the case of the Yiggy was that either method would increase quant noise, rounding would result a potentially less quant error and they decided it was low enough in level not to address it in by using a dithering filter. Adding dither would have added cost and complexity to the design and did so knowingly it would not measure in an ideal manner. In the test, it had shown that the rounded method used produced the exact result as truncating. No surprise in that part as rounding can produce the exact same figure. Rounding will in some cases, produce a slightly different result over truncation and have an very minor difference in quant error in comparison to truncation.

In large, this won't present much of an issue to most of recorded music. To the ones that it may have, will be quite low in level.