Measurements and Review of Schiit Yggdrasil DAC

[Frank Dernie]

If that's the case when why are they up in arms over it? If it's just mere interpretation you would think their reactions would be different, but the guys over there are going a bit bonkers over the results as if to say that Amirs measurements were either outright wrong or fudged.

As I think has been mentioned before, good reviews rarely are challenged but when fault is found owners of a popular product are likely to start attacking the messenger.

 

[Blumlein 88]

There are many sources of spurious signals that can appear in a dac output.

Jitter - This is a variation in the timing of the clock signal that the dac chip uses. In simple terms If the clock speeds up and slows down this directly changes the frequency of the signal the dac is producing. So if we play a fixed tone, such as the 12kHz in Amirs plots, you can see the frequency of the variation as discrete signals either side of the fundamental tone. The clock signal can also be affected by more random like noise and this appears as skirts at the bottom of the signal. AS Amir said the clock signal can get polluted by many sources including mains 50 or 60 Hz. You will usually see harmonic multiples of these frequencies, 100,150hz and 120, 180Hz. You often see this at those frequencies and as sidebands either side subtracted and added to the fundamental tone you are measuring.
Also the source of the clock has an impact. Asynch usb uses a clock signal generated by a component in the dac close to the dac chip. This minimises problems. Spdif and aes carry the clock from the source and it is extracted by the receiving dac using a circuit called a phase locked loop. Variations in the source and limitations of the PLL can allow jitter components to appear. Also data patterns can creaate jitter. Whilst many dacs are extremely good at extracting the clock from spdif/aes, some are not. USB avoids these issues.

Harmonic distortion - Any non linearity in the dac chip or the following output amplifier can cause muliples of the fundamental frequency to appear, 2,3,4,5 x etc.

Intermodulation distortion - related to harmonic distortion, again if non linear, where there are two different tones they will create further tones at frequencies of the two signals added and subtracted from each other. Several types of test show this, you will typically see 19 and 20kHz tones producing a 1kHz tone and 60Hz and 7kHz tones producing modulated sidebands of 60Hz around the 7 kHz tone.

Of course mains pick up is insidious and you can see 50 and 60 Hz plus harmonics on some dac outputs.

Other spurious signals can include idle tones produced in the dac chip and tones picked up from the general digital noise in the dac. Amir had already mentioned poor dac reference voltages and noise.

So really we are looking for anything that shouldnt be there, the above is a very general guide to the most obvious

I will add some pictures to this post later to help clarify.

Please view this thread and comment upon it.

 

[mindbomb]

@amirm

@Candlesticks mentioned this before as well, but isn't the low jitter on the yggdrasil largely due to the gen 5 usb upgrade? And doesn't atomicbob's measurements actually predate that being available? So is his measured jitter too good if it is at parity with yours?

 

[Randem Tox]

And doesn't atomicbob's measurements actually predate that being available?

His Yggdrasil v1 measurements would pre-date the USB Gen 5 board. His v2 measurements are on a unit with the USB Gen 5 board installed.

 

[Jimster480]

His Yggdrasil v1 measurements would pre-date the USB Gen 5 board. His v2 measurements are on a unit with the USB Gen 5 board installed.

Yes but as it shows, the Gen 5 board doesn't really improve much other than finally clean up the USB input to levels that other DAC's already have.

 

[garbulky]

I wrote a complete response to this in Mike's thread and it was this reply that was deleted and I was given a reply ban. So I say it is safe to say what I had to say was not to some people's liking.

I did not save the reply so here is another version.

The motivation for both running the linearity test and determining bit depths from it actually came from tests that Stereophile used to run. Here is an example: https://www.stereophile.com/digitalprocessors/187/index.html

View attachment 10764

If I gave you this graph, and I asked you where there is too much error, you would probably draw a line a bit after the curve gets wiggly, say, around -110 dB. That is exactly what I am doing with my 0.1 dB variation from linearity.

While linearity is shown as a figure related to input value (X axis), it is very common to instead talk about it in the order of bits. See: https://en.wikipedia.org/wiki/Differential_nonlinearity

And much more deeper treatment from the manufacturer of the DAC chip used in Schiit Yggy: http://www.analog.com/media/en/trai...ndbooks/Data-Conversion-Handbook/Chapter5.pdf

View attachment 10766

LSB is least significant bit. The error can be 1, 2, 3 or more of LSB, reducing the effective resolution of the DAC by the same number of bits. So we are talking about bits in linearity measurements.

Here is a partial spec from Analog Devices on this front:

View attachment 10767

As you see they show 20 bits as the maximum resolution of the DAC, but then show under different conditions how much that is degraded in units of LSB (bits). In the example I have shown the error can be +- 1 dB or 2 dB total making the effective, error free resolution 18 bits. Worst case actually brings the resolution way down to 15 (-2.5 to + 2.5 bits). Don't run with these numbers as these are just the example from above. I am showing them just to demonstrate the common interpretation of linearity in the form of bits.

It is also true that we can also convert signal to noise ratio into bits. Here ENOB (effective number of bits) is a simple way to express something as is how it is used above. Here we are trying to determine how noisy the DAC is. In linearity we are trying to determine how distorted it is.

As to Schiit spec, this is all they say:

View attachment 10765

What don't know what frequency was used for this measurement. Or whether they used weighting which filters a lot of the noise making the number a lot higher. It is also possible this is measured with no signal, and then a full amplitude signal. A proper signal to ratio would be looking at the noise in presence of our full amplitude signal as the noise level can increase with signal. Without this kind of detail or a spectrum graph as I post, there is just no way of knowing what this number is.

SNR is such an important marketing number that manufacturers resort to every trick in the book to come up with the best number they can.

Summary
ENOB or effective number of bits can be applied to both distortion and noise measurements. My use of it with respect to linearity is absolutely consistent with how DACs are rated and measured. It is however not a common practice or term in audio reviews and hence lack of familiarity with it even from that Phd engineer.

Ultimately you can ignore my ENOB computation and just look at the graphs. The visual completely makes the same point when the DAC is compared to another showing much less error.

So I've been mulling this over some. And I do have a few questions. But first, this is what bought up those questions. This is from the SBAF thread (where they've been bashing you).
Here they question the choosing of +/- 0.1db and how that would be used to equate number of bits.

I don't feel like I'm an Amir hater that much myself either.

But it is obvious to me that he does not have full understanding of what he is doing, and he is falsely projecting himself as if he does. This is not good.

He is grossly incorrect in his interpretation of his linearity measurements. The +/-1 LSB implies that linearity is bounded by +/- 6 dB, not +/- 0.1 dB. And again, 10-bit my ass. Moreover, many of his other measurements do not match @atomicbob's own measurements. Grossly so. His interpretation of a basic THD+N plot is wrong as well. This has to be called.

Bimby does not measure badly. It measures respectably for 16-bit entry level multi-bit DAC. And the fact that Amir paints the opposite picture incorrectly, with folks actually buying that, is only the tip of the iceberg. There is also the rest of the equipment he has reviewed.

↑
- Yeah.
What we can say is that he measured linearity that was accurate to 10 bits. (linearity fell off at -58 dB).

That is apparently not the same thing as ENOB. It is a different measurement.

So that just leaves the question, how important is linearity? Or, how how important is linearity past 10 or so bits?

Edit: or past 12 bits, if atomicbob's bimby is the normal one.

-I don't think one can say that he measured linearity that was accurate to 10 bits.

I think the y-axis in the linearity measurement corresponds to error (LSB or dB).

I think +/- 6 dB corresponds to +/- 1 LSB error.

I think Amir's +/-0.1 dB criteria corresponds to +/- 0.0166 LSB error.

LSB_error = log(10^((dB_error)/20))/log(2)

So Amir is saying that if the DAC has a +/- 0.0166 LSB error then the DAC has failed linearity at that point, and the number of "linearity bits" (whatever that means), is the level at which the DAC failed the +/-0.0166 LSB error. That makes little sense to me.

That is indeed not ENOB. That appears to be some made up number, based on some logic I have yet to understand. So what if I set the linearity error threshold to +/- 0.58518 bits, or +/- 3.1416 bits, or +/- 0.69 bits... what does that mean?

In Amir's linearity plot, if the linearity error goes over +/-0.1 dB then he claims arbitrarily that the system failed. See his section on "Linearity Test" here: https://audiosciencereview.com/forum/index.php?threads/understanding-audio-measurements.2351/

If I was to go by that then the Bifrost "fails" at -58 dB (10-bit). If I go by +/- 6 dB then we cross around -85 dB which is not 10-bit, more like 14-bits. Note, we cross that +/-6 dB around -85 dB in @atomicbob's plots as well. The y-axis are scaled differently.

Furthermore, that only says that we should expect about 1 LSB error when the signal is at -85 dB from full scale. Nothing more. Not 10-bit DAC or 24-bit DAC or whatever. Like you said, that's not ENOB.

So what does 10-bit DAC mean in Amir's world? The point at which the linearity measurement yields over +/- 0.1 dB of error? because it's not ENOB. That's misleading, don't you think?

(Atomicbob)
It is this type of misleading interpretation that prompted me to write the Lies ... DAC Tech Meas tutorial. You are quite correct in pointing out that the 0.1 dB threshold for gain non-linearity was pulled out of a very dark orifice. And also for pointing out that gain linearity has nothing to do with ENOB. They measure different converter performance attributes.

Another thing to ponder is the comparison of DAC architectures. Sigma Delta is practically gain linear by definition. R2R is not. But then that is obvious to Real Engineers(TM). Either the self-appointed measurement guru is not well informed, or he is and deliberately misleading. From the TI reference I gave earlier:

Delta Sigma—(One bit) Trades resolution in amplitude for resolution in time. Requires a system clock that is faster than the bit data

First I apologize for the rudeness in the quotes. I thought it was more honest not to edit the quoted posts. But feel free to do so.
- Do you have any response to the usage of 0.1 db as the cut off for ENOB vs 6 db for the LSB? - - Do you have any supporting evidence for using 0.1 db as the cuttoff. Also why don't you include a 6db LSB intepretation.
- How about the context of R2r architecture vs DS? Do you plan to do impulse measurements for the "time performance of R2r"?
- Also if SNR is the common method of determining ENOB. Why not include that measurement? And since the two intepretations of ENOB in SNR and the linearity graph appear to differ ....well it can't be one or the other right? And you mentioned that they will be difference because linearity measures linearity. Do you have evidence to show how to convert linearity to ENOB? Or are you measuring something different when you mention bit depth?

 

[amirm]

I am in the process of posting a review so I will give a brief answer. There are DACs that are linear up to 120 dB and varying no more than 0.3 dB or so. If we used 6 dB number they suggest, all DACs whether they are off by 5 dB or 0.3 dB would get the same resolution. Clearly this is not correct and would be an attempt to make a poor performing DAC look as good as others in this metric.

The other way to look at my metric is at which point/level a DAC falls from near perfection. You can then stick with dB numbers if number of bits bother you. But otherwise the number of bits is simply that dB number divided by 6.

BTW, what I am doing is not anything your eyes and lay intuition would do differently. Here are the linearities of Schiit BiFrost Multibit:

Looking at the graph on the left, wouldn't you pick a point close to where I have the X instead of going to -6 dB that is off the chart at the bottom (graph only goes to -5 dB).

BTW, the graph axis is what Audio Precision picked for this measurement. As you see, it only goes to +- 5 dB. That should tell you what they think they worst case measurements should be.

 

[garbulky]

I am in the process of posting a review so I will give a brief answer. There are DACs that are linear up to 120 dB and varying no more than 0.3 dB or so. If we used 6 dB number they suggest, all DACs whether they are off by 5 dB or 0.3 dB would get the same resolution. Clearly this is not correct and would be an attempt to make a poor performing DAC look as good as others in this metric.

The other way to look at my metric is at which point/level a DAC falls from near perfection. You can then stick with dB numbers if number of bits bother you. But otherwise the number of bits is simply that dB number divided by 6.

BTW, what I am doing is not anything your eyes and lay intuition would do differently. Here are the linearities of Schiit BiFrost Multibit:

Looking at the graph on the left, wouldn't you pick a point close to where I have the X instead of going to -6 dB that is off the chart at the bottom (graph only goes to -5 dB).

BTW, the graph axis is what Audio Precision picked for this measurement. As you see, it only goes to +- 5 dB. That should tell you what they think they worst case measurements should be.

Well it's not about what I would pick. Or what obviously looks better. Do you have any links or evidence to using +/- 0.1 db to determining ENOB? Or did you pick it because it looks better on the eyes?

Whether other DACs that deviate much much less will now get lumped in to the same category still doesn't change that there needs to be an accepted procedure for doing ENOB. You mentioned linearity can be used before. I want to know how and where 0.1 db comes in to it. And the reason for it is because that 0.1 db number appears to be what's causing drastically different numbers on bit depth away from the manufacturers spec.

Number of db doesn't bother me as much. And I'll accept that they appear to vary much more compared to the other DS DACs. What bothers me is that your intepretation of these measurements imply that the manufactuer is outright lying on their specifications of ENOB. And the usage of 0.1db is used to say or imply that this is indeed the ENOB of the DAC.
If so, I think a documented accepted procedure of intepretation (for the 0.1 db) needs to be shown for ENOB.

So another way of saying it could be more all these dacs have very similar equivanlent number of bits. But look at the variation in the last 1/8th or what have you of the LSB. Which seems more accurate than "the 21 bit yggy is a 16 bit dac"!

I know you are busy. When you have a chance I would appreciate your thoughts on my last two questions on the previous posts regarding impulse response, and posting SNR data.

 

[Blumlein 88]

Well it's not about what I would pick. Or what obviously looks better. Do you have any links or evidence to using +/- 0.1 db to determining ENOB? Or did you pick it because it looks better on the eyes?

Whether other DACs that deviate much much less will now get lumped in to the same category still doesn't change that there needs to be an accepted procedure for doing ENOB. You mentioned linearity can be used before. I want to know how and where 0.1 db comes in to it. And the reason for it is because that 0.1 db number appears to be what's causing drastically different numbers on bit depth away from the manufacturers spec.

Number of db doesn't bother me as much. And I'll accept that they appear to vary much more compared to the other DS DACs. What bothers me is that your intepretation of these measurements imply that the manufactuer is outright lying on their specifications of ENOB. And the usage of 0.1db is used to say or imply that this is indeed the ENOB of the DAC.
If so, I think a documented accepted procedure of intepretation (for the 0.1 db) needs to be shown for ENOB.

So another way of saying it could be more all these dacs have very similar equivanlent number of bits. But look at the variation in the last 1/8th or what have you of the LSB. Which seems more accurate than "the 21 bit yggy is a 16 bit dac"!

I know you are busy. When you have a chance I would appreciate your thoughts on my last two questions on the previous posts regarding impulse response, and posting SNR data.

Well ENOB is usually not in relation to low level linearity.

http://www.analog.com/media/en/training-seminars/tutorials/MT-003.pdf

Look down on page 5.

ENOB is signal to noise and distortion. ENOB equal max signal ratio to noise and distortion levels with N= 6db per bit - 1.76 db.

So ENOB is not quite what you are really looking for here.

That doesn't mean Amir's measurements are wrong, and what keeps being overlooked is Atomic Bob's measurements show very much the same thing. So the measurements concur. The conclusion drawn differs. Personally I wouldn't use low level linearity to define bits of resolution. The Schiit device has poor linearity for a multi-bit DAC. As I posted on the other forum, and got kicked off the thread for it: I can't find measurements of multi-bit DACs as poor as the Schiit. The worst I found was for a 47 Labs DAC and it was better than the Schiit.

Multi-bit DACs don't have linearity equal to sigma delta devices. If the linearity is pretty good it probably is of no audible consequence. The Schiit result however is egregiously poor. Do we overlook one of the worst results among multi-bit DACs in a $2000 device? It doesn't make sense to me.

I haven't heard the Yggy, does it sound great? Maybe it has a sound, maybe that is a nice sound. Is it due to this linearity or something else? For that kind of money I would spring for a Benchmark with superlative performance and do my pleasing colorations somewhere else in the playback chain. Or if I want multi-bit the Soekris reviewed elsewhere on ASR looks like a much better device for less money. It gets to 17 or nearly 18 bits by Amir's criteria.

 

[garbulky]

Well ENOB is usually not in relation to low level linearity.

http://www.analog.com/media/en/training-seminars/tutorials/MT-003.pdf

Look down on page 5.

ENOB is signal to noise and distortion. ENOB equal max signal ratio to noise and distortion levels with N= 6db per bit - 1.76 db.

So ENOB is not quite what you are really looking for here.

That doesn't mean Amir's measurements are wrong, and what keeps being overlooked is Atomic Bob's measurements show very much the same thing. So the measurements concur. The conclusion drawn differs. Personally I wouldn't use low level linearity to define bits of resolution. The Schiit device has poor linearity for a multi-bit DAC. As I posted on the other forum, and got kicked off the thread for it: I can't find measurements of multi-bit DACs as poor as the Schiit. The worst I found was for a 47 Labs DAC and it was better than the Schiit.

Multi-bit DACs don't have linearity equal to sigma delta devices. If the linearity is pretty good it probably is of no audible consequence. The Schiit result however is egregiously poor. Do we overlook one of the worst results among multi-bit DACs in a $2000 device? It doesn't make sense to me.

I haven't heard the Yggy, does it sound great? Maybe it has a sound, maybe that is a nice sound. Is it due to this linearity or something else? For that kind of money I would spring for a Benchmark with superlative performance and do my pleasing colorations somewhere else in the playback chain. Or if I want multi-bit the Soekris reviewed elsewhere on ASR looks like a much better device for less money. It gets to 17 or nearly 18 bits by Amir's criteria.

Thanks for your response. First I tried to read the pdf but I'm not sure exactly what you wanted me to see. I didn't see anything about linearity there? But I did see the SNR calculation.
I absolutely agree that the less than stellar linearity should be mentioned. It's important information. No arguments. My question is the statement of bits and how it was obtained. If there is no supporting evidence for using 0.1 db coupled with linearity to call a DAC 16 bits for the Yggy or 10 bits for the bifrost then it should be restated using proper procedure.

I see no reason to take out linearity measurements.

 

[Blumlein 88]

Thanks for your response. First I tried to read the pdf but I'm not sure exactly what you wanted me to see. I didn't see anything about linearity there? But I did see the SNR calculation.
I absolutely agree that the less than stellar linearity should be mentioned. It's important information. No arguments. My question is the statement of bits and how it was obtained. If there is no supporting evidence for using 0.1 db coupled with linearity to call a DAC 16 bits for the Yggy or 10 bits for the bifrost then it should be restated using proper procedure.

I see no reason to take out linearity measurements.

I'm not aware of others using linearity in terms of bits this way. Closest I have run across is dCS which uses an error of 1 db instead of .1 db as something of a cutoff point for accurate low level linearity.

I think people are reacting to the text on Amir's graph pointing to 10 bits or 16 bits. Maybe it should be called lowest accurate bits. That fits what it is. It isn't raw bits of resolution nor is it really ENOB. Lowest accurate bit is what it is. I also would use 1 db with such low signal levels, but that is up to Amir. Anyone can view his graphs and see where that lies if they wish. In the above example 1 db would occur just past 12 bits which is still very poor. The Topping DX7 goes to the bottom of the chart with less than .5 db error at the 20 bit point. The 47 Labs which is the worst I recall JA at Stereophile measuring. Was off by 2 db at -100 dbFS.

http://sdg-master.com/lesestoff/bits.pdf The guys at dCS discuss linearity. I think their claim of 24 bit linearity is a bit much considering their results, but everyone wants their gear to look good.

 

[Dimitri]

Maybe it's just me, but questions from other forums (i.e. from Small Beaks Audio Freaks ) should be freely plagiarized, copied with no credit , recognition or reference to the source.

If 'ya have a question, ask a question.

But this game of "here is what they said. What's your response" Is about to start getting old and annoying to me. I could be the only one.
I could be wrong. I was once.

 

[Dimitri]

The Schiit result however is egregiously poor. Do we overlook one of the worst results among multi-bit DACs in a $2000 device? It doesn't make sense to me.

You are supposed to sweep it under the rug as if it weren't there. Therefore, nothing to overlook.It wurks gud.

 

[DonH56]

ENOB is a dynamic measurement set by SINAD.
The linearity plots are essentially DC (static) linearity.
Two different things.

Using 0.1 dB for deviation from linearity in the linearity plots is consistent across Amir's data and is useful as a relative basis for comparison among various DACs. The plots are there if you want to use a different metric but arguing about what is correct is sort of pointless. You can see the differences. Whether or not they matter is a better topic for debate IMO.

 

[Sal1950]

I could be wrong. I was once.

I thought I was wrong once,
Till I realized I was only mis-taken.

 

[Thomas savage]

Maybe it's just me, but questions from other forums (i.e. from Small Beaks Audio Freaks ) should be freely plagiarized, copied with no credit , recognition or reference to the source.

If 'ya have a question, ask a question.

But this game of "here is what they said. What's your response" Is about to start getting old and annoying to me. I could be the only one.
I could be wrong. I was once.

No, I agree.

 

[mindbomb]

@garbulky
I think I can help you understand this. A dac's bitdepth can describe many things. Sabre dacs are sometimes referred to as 32 bit because they accept 32 bit data. This is not that relevant for consumers, but that is one way to do it. Another thing bitdepth can refer to is the dynamic range of the dac, which is a reflection of noise performance. This is represented visually here by the -90db sine wave. The better the sine wave looks, the more dynamic range it has, the better the noise performance. Then there is linearity also, which checks to see if the dac delivers the right voltage at different levels. So for example, if you had a dac with really poor linearity, the quiet parts of a song will end up playing at the wrong volume. So I described 3 different ways to talk about the bitdepth of a dac, and they can all be different values. For example, with the topping d30, it can accept 24 bit data, it has 19 bits of dynamic range, and 15 bits of linearity within 0.1dB.

On top of that, then there is the audiophile way to define resolution to make r2r dacs look really good - which is maximum theoretical dynamic range without oversampling or noise shaping. This is how it is done on the schiit website to explain how the yggdrasil is 21 bits and other dacs are 6 bits. To me, this is extremely contrived.

On sbaf, it sounds like they are cynically seizing on the fact that 0.1dB is intuitively a very small value to sow doubt in the measurements. Yes, it might be a hard value to hear, but at the same time, from a technical perspective, it is very unusual for the bifrost multibit to have these problems as early as it does. The graph is there so you can choose your own error threshold if you wish. That doesn't change the fact that its linearity is poor relative to other dacs.

 

[Jimster480]

Thanks for your response. First I tried to read the pdf but I'm not sure exactly what you wanted me to see. I didn't see anything about linearity there? But I did see the SNR calculation.
I absolutely agree that the less than stellar linearity should be mentioned. It's important information. No arguments. My question is the statement of bits and how it was obtained. If there is no supporting evidence for using 0.1 db coupled with linearity to call a DAC 16 bits for the Yggy or 10 bits for the bifrost then it should be restated using proper procedure.

I see no reason to take out linearity measurements.

What they are trying to explain to you is that the measurement is so bad that even if they used 0.3 db or even 1db it doesn't come anywhere CLOSE to the claimed measurements.
The number of bits that a DAC is only relates the number of bits it takes for input values or the number of bits it uses internally to process.

If you read the schiit marketing they make every DS DAC sound horrible and that you lose all of the bits of resolution in your sound. But when measured, their Multibit DAC's actually do exactly what they are fauulting DS DAC's for.
The measurements are so poor that many $1 TI DAC chips will outperform most of their lineup, if not the entire lineup.
IF you want to go to a "flagship" Burr-Brown (TI) DS DAC..... it will mop the floor with Yggdrasil too with the right implementation and the chip costs $12.
Simply put... Schiit is marketing, their goal is to sow doubt in other products to get you to buy into their tiered listening scheme where they make each tier sound audibly better than the last tier so you can feel like you are getting your moneys worth. But in reality their best performance is shy of what other products offer for $100-200.....

They can try to sow doubt in Amirs measurements, but the measurements are literally so bad that its not an argument of which db's of error you would have because they still fail unless you go outside of their db's (in this case 6+) which as Amir said is outside of the graph chosen by audio precision themselves. And considering that Audio Precision is a company that is specifically devoted to measurement equipment and the standard of what is well performing or not performing..... their suggestion has much more weight than any audio mfg especially a singular mfg that happens to have worse measurements than all other mfg's...

 

[RayDunzl]

@amirm

Is the linearity test performed with or without dither?

How would the results compare between the two choices?

 

[RayDunzl]

Sabre dacs are sometimes referred to as 32 bit because they accept 32 bit data.

I'd think the 32bit nomenclature comes because of internal conversion (selectable?) to 32bit before further internal processing..

My DAC doesn't accept 32 bit, but the digital volume function uses them.

However...

So... OK. Time marches on...