Schitt's flagship DAC uses a not-for-audio DAC chip.

The Schitt Yggy DAC has already created something of a stir over on Head-Fi amongst those who've heard pre-production prototypes. Its of interest not just because of the pre-launch buzz but also because the designer (Mike Moffat) is one of the long-time seasoned guys in the field. He says this is the best practical DAC he knows how to build. And its priced ISTM very reasonably given the amount of tech it embodies ($2300) - the DAC chips come from ADI and are $64 a piece on the manu's website. He's using 4 of them but says he had to address the glitching of the DAC without using a sample-hold which sounds like ***.

The chip is the AD5791BRUZ which ADI designed for industrial/scientific/medical applications rather than for audio. Datasheet attached.

Having looked over the DS what strikes me as interesting is that this is a 20bit DAC (1ppm) yet the 1kHz THD performance (p4) isn't anything to write home about (-97dB) and that figure is given at a very low sample rate (10kHz). What this suggests to me is that performance is 20bits at DC and very low frequencies, but is degrading quite rapidly within the audio band. Perhaps this is the reason that Mr Moffat had to take steps to deglitch it. From what I recall, THD (not THD+N) theoretical performance is about 9dB per bit for a perfect DAC - hence the THD performance here is only that of a perfect 11bit DAC.

Compare this measurement with the TDA1541A which delivers -100dB THD @1kHz but at a sample rate of 176kHz (nearly 18X that of the ADI chip).

Also worthy of note is that the headline settling time (1uS) is spec'd out in two ways. A major-size step but to 0.02% (around 12-13bits) and a 500step code change (equivalent to an LSB of an 11bit DAC). Using this DAC at the maximum sample rate a power of two multiple of 44.1kHz would give 768kHz (16X OS) so the biggest step would be about 1/16th of FS. I can't help thinking that the exemplary DC specifications are almost entirely irrelevant when the output's spending such a large proportion of its time settling to the next value.

Update - the piece is now shipping out. I still haven't got a cogent explanation of 'closed form digital filter' and why that would mean all original samples are preserved. Not for want of looking around though. If anyone's getting an Yggy I'd love it if close-up photos of the DAC boards were posted somewhere so I could get an appreciation of how Mr M's implemented his deglitcher. It would be very interesting to see if the '21 bit resolution' claim really stands up in practice and the deglitcher's dynamic performance is key to this.

Update2 - technical measurements are now up on Head-Fi : https://www.head-fi.org/t/764787/yggd...l-measurements

Something is puzzling me about the THD vs THD+N plots (4th graph down the page). The figures are screaming 'S-D' at me because nowhere else have I seen THD+N figures so much worse than THD except on some S-D DAC datasheets. What's even more concerning is the rise of both at 20Hz to a level around 0.1% (-57dB). I'd expect a rise in THD (but not THD+N) if there was a coupling trafo in there. So where could all that noise be coming from and what form is it taking? THD+N at 0.01% when THD is 0.0024% is really weird and is looking distinctly like noise modulation. Any guesses - are we seeing something the deglitcher is contributing or is this the 'closed form' digital filter?

Update3 - comparison review of Yggy against a Benchmark DAC/PRE - https://www.head-fi.org/products/schi.../reviews/13419
The comments in this review make a lot of sense to me, outlining the differences commonly encountered between 'designed by the numbers' DACs and those designed for listening pleasure. I've been aiming to get maximum soundstage depth from my own DAC designs and its interesting to me that this is one of Yggy's strong points. In my experience, good soundstage depth correlates with not having the very lowest level (and lowest frequencies) signals corrupted by signal-correlated noise. In contrast the Benchmark sounds more forward - meaning its adding in signal-correlated distortions which the ear/brain can't place into the acoustic space. I'd hazard that connecting the Benchmark to just a single speaker via a transparent amp would reveal that the surface of the tweeter was seeming to produce sound, rather than being a window to sounds coming from further behind the dome.

Update4 - just noticed that the multibit retrofit for the next lower priced Schitt DAC is becoming available. Seems they're using the 18bit variant of the 20bit DAC in the Yggy.

Update5 - here are some more measurements of the Yggy done with QA400. https://www.changstar.com/index.php/topic,2772.0.html

The discussion towards the end about 120dB SNR is based on a misunderstanding about how to read FFTs. SNRs in audio are normally within the whole audio band - measuring noise without a bandwidth is meaningless. An FFT's measurement bandwidth is normally a very small fraction of the audio band - in those FFTs the bandwidth is shown, its 11.7Hz in the widest bandwidth plots, 2.9Hz in others and 0.37Hz in the plot with the highest level of 'zoom' (yggy3).
So when pulling a dB number from an FFT do take care to notice the bandwidth associated with that number.

Worked example - on yggy3 the LF noise starts to rise up from the baseline. At 40Hz we see about -120dB in the 0.37Hz bandwidth. If the DAC put out that level of noise but in the whole 20kHz audio bandwidth we'd see a decrease in the dB number by 10log(20k/0.37). I calculate that increase as 47dB, meaning the LF noise 'floor' corresponds to 120-47 = 73dB SNR. Mid-band the figure's around 136dB which means an AP should show the audio band SNR as 89dB. Note that the DAC's only receiving 16bit data which should have dither noise around -93dB so its looking like this DAC is noisier than 16bits.

Oh and just in - Chris C over at Computer Audiophile just now got his hands on an Yggy - 'holy schitt' - https://www.computeraudiophile.com/f6...-week-~-25657/

Go here to see some guy doing internal mods - plenty of pics of the PCBs I can almost read the part numbers - https://www.tweakaudio.com/EVS-2/Schiit_mod.html

Update6 - some more light (perhaps) shed on the nature of the digital filter, my nagging question about 'closed form filter' still awaits a clear answer. Mike has signed up to CA and posted here : https://www.computeraudiophile.com/f6...tml#post463821

Let's dissect what he's saying. Firstly

It keeps all original samples; those samples contain frequency and phase information which can be optimized not only in the time domain but in the frequency domain. We do precisely this; the mechanic is we add 7 new optimized samples between the original ones.

I'll need to do some research on this but to me (as far as I can recall) this makes it a half-band filter by keeping all the original samples.

The common digital filter method is a Parks-McClellan algorithm, which has been used in all of the older oversampling chipsets, and persists to this day as the input filter in most ds DACs.

Here his writing isn't totally clear so let me unpack it somewhat - he's talking about how filters are designed, mathematically. Parks-McClellan is a way (an iterative algorithm) for getting the best optimized coefficient set given the desired parameters (ripple, stop band rejection). He's making an assumption here - he doesn't know for sure that all digital filters were designed this way but its a fairly reasonable assumption. If you design your filter with Matlab then it'll use Parks-McClellan (I think) to get the coefficient set.

This of course, means it never completely solves. The worse news is that all original sample are lost, replaced by 8 new approximated ones. Further, the Parks McClellan optimization is based on the frequency domain only – flat frequency response, with the time (read spatial) domain ignored.

This 'never completely solves' is misleading. Since the results are by necessity quantized coefficients then precise solutions are quite useless. No matter how the filter's designed, quantization of the coefficients is a given.The second sentence also is misleading - a half-band filter can be designed using Parks-McClellan and hence it can keep all the original samples. Saying 'approximated' is a third example of being misleading - in his filter since only 1 in 8 of the output samples is original it must follow that the other 7 are 'approximated'. By his own spin his DAC is putting out 87.5% approximated samples. Lastly the P-M algorithm can design linear phase filters so the time domain isn't ignored necessarily.

Our filter is based upon closed form math – the coefficients are not approximations, the equations solve; the matrices invert and the math is done. The filter also optimizes the time domain.

Here he sounds to me to be saying something quite radical. If indeed the coefficients coming out of his closed form math are precise ones, it means they have no need to be quantized. Which would mean that the closed form math also tells the filter how long (in terms of bits) each coefficient must be. He's saying his math produces precise integer solutions for all the coefficients. I'm most certainly not a mathematician but it sounds to my ears a far-fetched claim to make. I'd really like to hear from a mathematician about this - any reading? From my childhood reading of Martin Gardner, looks like this gets us into the realm of 'diophantine' equations. It wouldn't map well to fixed form digital filter implementations since having the math tell us how many bits we need for our coefficients is inconvenient - we need to design the filter before we can design the hardware.

For now I'll interpret Mike's writing that when he says 'coefficients aren't approximations' he's excluding quantization in his meaning of 'approximation'. Which then means his 'closed form' filter design method appears to have nothing to recommend it over Parks-McClellan.

Update7 - Bifrost now has a multibit version available and its using an AD5547 - a dual 16bit parallel input, current output multiplying DAC. Datasheet attached.

Mike's putting out some misinformation about the AD5547 used in the 'Bimby'. I noticed this on CA, which presumably originally comes from HeadFi -

The THD performance is actually scary good--far better than 16 bits. 16 bit level THD, from a theoretically perfect 16-bit DAC, is -96dB.

Nope, he's confused THD with THD+N here. The 'N' component of this measurement means the best you'll get from 16bits is about -93dB THD+N (assuming an original analog waveform was dithered at the ADC) but if you generate the waveform digitally and feed that to a 16bit DAC you can get a better figure just for the distortion, such that the theoretical maximum is about 9dB per bit, so -144dB for THD (excluding N). However no DAC in existence has such a good noise floor so it'll be noise dominated in practice.

The AD5547 is -104dB, much better than 16 bit-

Thereby he contradicts himself - the AD5547 is better than the theoretical best?

Its worth pointing out here that the DS spec for the THD+N for this DAC isn't at all impressive, rather like that for the 20bit part in the Yggy. -85dB @ 1kHz, this time without mentioning the sample rate. The glitch also is fairly high at 7nV-s so its not a chip you'll want to run fast. Some of TI's non-audio DACs do at least 10X better for glitches.

Also worth noting that the advertised settling time of 0.5uS is only for 0.1% precision (i.e. 10bits) using an AD841 as I/V.

Update8 - there's been some criticism of Yggy from Peter and Mani (Peter's the designer of the multiple PCM1704 Phasure DAC, Mani's one of his delighted customers) over on CA with reference to a scope plot allegedly showing 'glitchiness' at -90dB. Mike M was hospitalized for a while so hadn't been able to respond, but now he has and his post you can find here : https://www.computeraudiophile.com/f6...tml#post541561

Don't go there if you don't like flowery language incidentally. In summary Mike's saying the glitch is real and its the zero crossing glitch common to all R2R converters - the MSB carry problem in other words. At zero, R2R DACs change all their switch positions but there's always some skew (time delay) between them so they don't all switch at the same instant, meaning during the transition all kinds of other codes get generated. Which rather begs the question - why doesn't the deglitcher handle this? There's been quite a lot of talk about how the AD DAC doesn't do its stuff alone, it needs the help of a deglitching circuit.

Update9 - Steve Guttenberg now has a review up for the Bifrost MB : https://www.cnet.com/news/the-good-sc...iophiliac.ftag
My one sentence summary is - 'this DAC has dynamics and tonal colours, in comparison the S-D version sounds 'greyed-out''. But you don't need to spend $600 to get dynamics and tonal colours, a TDA1387 based DAC will give you those, in an optimized implementation, for much less money : https://www.diyaudio.com/forums/blogs...ding-base.html
I do rather disagree with his recommendation to 'invest in the best speakers and amp you can before upgrading your source'. Optimizing a system is rather more complex than just where the money's allocated but 'garbage in, garbage out' (Linn's marketing philosophy going way back) can't be negated. Components downstream of the source cannot 'add goodness' they can only subtract from it.

Update10 - they've now announced the Modi multibit, a total snip at $249 : https://www.computeraudiophile.com/f6...-%24249-29297/

Update 11 - some listening comparison details between Yggy and the Kitsune Holo Spring DAC. First detailed remarks I've come across which identify specific flaws in its sound : https://www.computeraudiophile.com/co....html#comments

Stereophile now has an Yggy review - https://www.stereophile.com/content/s...OQOrLet3WSB.97

I've been following Jason's blog over on HeadFi for quite some time now. He's just posted up a new entry with a fascinating peek behind the scenes of US-based semiconductor pricing. I'm betting that the 'expensive' chips he's talking about here (scroll down to 'The Deck, and How its stacked') are the Yggy's DACs from ADI.

https://www.head-fi.org/f/threads/sc...#post-13467654

Attached Files

	AD5791.pdf (1.12 MB, 650 views)
	AD5547.pdf (314.5 KB, 310 views)