DAC audible differences are finally proven with ABX?

[pma]

The line level part is easy enough to design if you are happy with the performance of the TI PGA2311 volume control (-107dB THD+N).

I assume you would probably want to get -115dB to -120dB THD+N. Separated grounds, constant output and input impedance, no crosstalk, no added hum components in the final setup. I can guarantee it is not easy to get. If you built or will build something similar, you will see.

 

[Thomas_A]

I think there aee a bunch of people here having a competent AD converter like the Cosmos. Now also having one DAC that Josh tested could make a DA/AD conversion and provide the file. With some luck we could have a tested pair for deltawave.

 

[KSTR]

The variable resistor is the load for the source/DAC (voltage divider). Its non-linearities directly distort the output of the DAC as evidenced by how the same happens with distortion created in dummy loads for amplifier testing. An active load is just not going to be clean and certainly not for state of the art DACs.

This is correct but not what the plot you posted shows.
It shows integral nonlinerity, that is the deviation from ideal when applying a constant input signal and stepping through the codes which set the point at which the resistor ladder is tapped. Note than in operation the code setting is not changed in the ABX box, just for clarity.

In case of rheostat mode, Fig.39 applies where we can see a constant current is injected and the wiper-to-GND resistance is measured (by measuring the wiper voltage).

In Rheostat mode the distortion with AC signal dominantly comes from the variable wiper resistance vs voltage because the wiper is an array of simple CMOS switches with the typical graph shown in Fig.17:

As we can see, the Rheostat INL exactly tracks the wiper resistance profile (vs. supply voltage) as it should be, because the output voltage varies. At the point where the additional voltage drop on the wiper is largest the INL is off the most in the positive range (higher resistance than should be)

We must note this is a graph taken with an old school method of a stepped signal applied while stepping the codes, so this is actually a re-used analog time-domain plot rather than a correct pin diagram. The input signal is set to a new value every 16 codes, as a part of the measurement procedure, it is not a behavior of the part itself, whose structure looks like this:

The silicon resistor string also has voltage coefficient (and temperature coefficient) but much lower.
All these distortions are relevant when the part actually was used in Rheostat mode which is what educated guessing seems to suggest, looking at the PCB.

In potentiometric mode, the distortion is much lower, notably when not loaded down. The wiper impedance is irrelevant, and the voltage coefficient of the divider with 256 all euqual resistors cancels out. That why in this mode AD states the distortion with 0.014%. Not stellar but good enough I would think to not spoil the test results.

As noted, we don't know which mode is used in the device but if Rheostat mode was used then that certainly would be an unlucky choice.

 

[ppierre]

Wouldn't we lose less dB by doing level matching in the digital domain before DAC.
And in the same KIS spirit, do fade out on the input signal to mask or simulate switching.
An LIRC remote for user input and you have enough for testing speakers.
For electronics devices the less obtrusive switching is still needed (controlled by the computer).

 

[KSTR]

I'd be happy to commission someone here to build *the ultimate ABX box* with better circuity and the addition of digital I/O. Also, silent switching. Take the current Van Alstine and let's beef it up to the point of future-proofing for the next decade. I'd gladly pay $2000-3000 for such a device. I'm dead serious.

OK, I'm going to think about this (a SOTA ABX hardware frontend for balanced and unblalanced line level signals, to be used with a computer -- via isolated USB-MIDI, for example, for control interface -- where the control software runs). But it would be a slow process as I still have a day job as electronic design engineer. I would provide stuffed and tested PCBs, to be built into some case.

What do you mean exactly with digital I/O? Do you need speaker level stuff?

 

[solderdude]

we don't know which mode is used in the device but if Rheostat mode was used then that certainly would be an unlucky choice.

Everything points to rheostat mode (looking at the PCB) but distortion does not appear to be as high as 0.014% given the measurements of @Josh83 as 0.014% would be close to audibility thresholds for single sine-wave signals rendering the device pretty useless.

On Multitone 32, inputs A and B returned IMD of 0.000307% and 0.000307%, TD+N of 0.000259% and 0.000258%

 

[KSTR]

Using an ADC for calibration and a switched resistor chain we can get to 0.02dB level matching between DUTs, _and_ also level match each DUT's right and left channels if you want, to eliminate that possible cue.

That's a critical point indeed. Precision A/B level matching requires better precision L/R channel matching to begin with, but that sort of alters the DAC when viewed as a black box. Same for polarity inversion.
So it is of utmost importance to fully state the goals and conditions. There is a big difference between testing DACs as is (black box) -- except some reasonable level matching -- and testing DACs with all tricks applied to make them sound the same as good as possible (modified black box) ... an option that a user won't have in practice.

 

[solderdude]

I think there are a bunch of people here having a competent AD converter like the Cosmos. Now also having one DAC that Josh tested could make a DA/AD conversion and provide the file. With some luck we could have a tested pair for deltawave.

I don't think/suppose that @Josh83 is likely to help invalidate his test by either posting his test files or recorded output. Because if that were to show some tells his show is over and needs to revise his article and redo it in a way some nerds at ASR would like it.
I hope I am wrong about this and at least posts the actual up/down sampled test files he used.
That would be a good place to start to investigate the anomalous findings.

From there it would probably be a good idea to record the output of each DAC (no ABX box yet) and see if there is clipping or ISO not handled correctly.
Step, by step.

That could potentially show why (if tells were present) how he could tell them apart which would at least validate his findings.

 

[Thomas_A]

I don't think/suppose that @Josh83 is likely to help invalidate his test by either posting his test files or recorded output. Because if that were to show some tells his show is over and needs to revise his article and redo it in a way some nerds at ASR would like it.

Yes, my suggestion would be to provide such files at ASR, just a DA/AD converted file of DACs used.

 

[amirm]

Wouldn't we lose less dB by doing level matching in the digital domain before DAC.

That's what I have done in the past, using Roon player to group DACs and then play the same content through them. I have then used a headphone amp with dual inputs to select between them.

The notion of routing the output of a state of the art DAC through a much lower quality analog device just doesn't sit well with me. When testing products like this, I listen to very low level signals that I then highly amplify. A comparator which adds noise and linearity errors totally messes with this.

Alternatively, start with two DACs with nearly identical output as to obviate the need for level matching. Plenty of DACs have nominal 2 volt output. 0.1 dB is about 25 millivolts so not too hard.

Heck, for the purposes of determining if the Author can pass such tests, we could even mod a DAC and adjust its output to match another if the above can't be found easily.

 

[mcdn]

I assume you would probably want to get -115dB to -120dB THD+N. Separated grounds, constant output and input impedance, no crosstalk, no added hum components in the final setup. I can guarantee it is not easy to get. If you built or will build something similar, you will see.

True, probably best to use relays to control opamp gain. @IVX could probably build this in something the size of a matchbox, it's basically a Cosmos Scaler with extra smarts

 

[NTTY]

On Multitone 32, inputs A and B returned IMD of 0.000307% and 0.000307%, TD+N of 0.000259% and 0.000258%, noise at −107.3 dB on both, and SNR of 104.9 dB on both.

Everything points to rheostat mode (looking at the PCB) but distortion does not appear to be as high as 0.014% given the measurements of @Josh83 as 0.014% would be close to audibility thresholds for single sine-wave signals rendering the device pretty useless.

Unfortunately, it is not possible to have a TD+N of 0.000259% (-111.7dB) with an SNR of 104.9dB. So there was an operator error for that measurement.

 

[RandomEar]

[...]
Are we missing some valuable data from only measuring in the frequency domain or have that been ruled out?

This is a bit of a discussion for a different topic. In the context of this thread, nothing much has been measured at all about the comparator, which is one of the points being criticized.

However, in general the frequency domain and the time domain are interchangeable. That's why you can do a Fast Fourier Transform (FFT) on your data, follow up with an inverse FFT and - apart from tiny errors due to numerical limitations - end up with an identical signal in the time domain.

But there's a but: FFTs do use window functions, so their time resolution is limited. And all the graphs we typically see are from relatively long, time averaged FFTs over seconds and sometimes maybe minutes. That increases the SNR, which is desired. However, if you assume that a device has a hypothetical small glitch right at the beginning of the measurement, that glitch would be "averaged out" the longer you capture data for your individual time-averaged measurement. That's just how it is and is technically not even related to the fact that the frequency domain is being analyzed.

In addition, you do need to see the phase response on top of the frequency response to get the full picture in the frequency domain. That is rarely shown here on ASR. I would assume that's because it is very uneventful most of the time. If you analyze in the time domain and just measure the impulse response, you'll get the full transfer function of the device in a single plot which is very convenient. I'm sure that some types of defects are easier to spot in one domain than in the other, so there is value in both paths of analyis.

But now back to proposed audible differences in DACs!

 

[mcdn]

So it is of utmost importance to fully state the goals and conditions. There is a big difference between testing DACs as is (black box) -- except some reasonable level matching -- and testing DACs with all tricks applied to make them sound the same as good as possible (modified black box) ... an option that a user won't have in practice.

It's important to be able to do both I think, because . But I can't think of anything except channel matching and output loading an ABX device would actually do. Any other ideas?

 

[mcdn]

The notion of routing the output of a state of the art DAC through a much lower quality analog device just doesn't sit well with me. When testing products like this, I listen to very low level signals that I then highly amplify. A comparator which adds noise and linearity errors totally messes with this.

Of course! this is all trivial when starting from a position that A/D and D/A conversion is transparent in the first place. But there's also a convenience factor that @hardisj has mentioned, of being able to choose A/B/X from the listening position easily

 

[welwynnick]

........ However, during R&D, the analog guys is constantly trying to verify data in the frequency domain and in the time domain to figure out issues. I am not an expert in analog signals - I stick to digital logic, but that makes me wonder: Are we missing some valuable data from only measuring in the frequency domain or have that been ruled out?

However, in general the frequency domain and the time domain are interchangeable. That's why you can do a Fast Fourier Transform (FFT) on your data, follow up with an inverse FFT and - apart from tiny errors due to numerical limitations - end up with an identical signal in the time domain.

This is a subject of particular interest to me, especially as time domain response is often brushed aside as simply being a convolution of the frequency domain.
It's exactly that, but only when the phase response is the same.

DAC tests often have impulse test charts, but they rarely have phase response. SoundstageNetwork are an exception, and they now include phase response in their DAC tests.

Phase response should be linear - like amplitude - but it rarely is. Phase response is different for each DAC, for each sample rate, and for each reconstruction filter.

I think that getting fixated on sky-high SINAD values on a high level 1kHz single tone isn't very instructive, and only leads one to think that everything must sound the same.

 

[Thomas_A]

It's important to be able to do both I think, because . But I can't think of anything except channel matching and output loading an ABX device would actually do. Any other ideas?

Still remains polarity. Devices have shown to be variable in this regard. Although this device seems to retain it. But DACs can vary.

 

[NTTY]

Alternatively, start with two DACs with nearly identical output as to obviate the need for level matching. Plenty of DACs have nominal 2 volt output. 0.1 dB is about 25 millivolts so not too hard.

Quick look at my inventory, I found two with 0.13dB to 0.15dB (left-right) differences (SMSL PS200 and PL200MKII).

 

[Roland68]

That's what I have done in the past, using Roon player to group DACs and then play the same content through them. I have then used a headphone amp with dual inputs to select between them.

The notion of routing the output of a state of the art DAC through a much lower quality analog device just doesn't sit well with me. When testing products like this, I listen to very low level signals that I then highly amplify. A comparator which adds noise and linearity errors totally messes with this.

Alternatively, start with two DACs with nearly identical output as to obviate the need for level matching. Plenty of DACs have nominal 2 volt output. 0.1 dB is about 25 millivolts so not too hard.

Heck, for the purposes of determining if the Author can pass such tests, we could even mod a DAC and adjust its output to match another if the above can't be found easily.

For our tests, we use only voltage dividers to adjust the volume or output voltage. The first stage handles coarse adjustment and the second fine adjustment, ensuring that only two high-quality resistors lie in the signal path—eliminating the need for relays at this point—while achieving precision to three decimal places. To this end, we utilize multiple sets of E192-series precision resistors that we have also selected in matched pairs.

 

[NTTY]

Thanks for sharing.
Now we know how the attenuation is done, using AD5262 digital pots. And we see loud monster relays, not really ideal for line-level switching. No coupling caps with pre-charging in the inputs. No GND side switching (which would ideally require active diff inputs). Sloppy wiring. And then we have those weird signal-combining RCA jacks with 10k(!) isolation resistors which I assume were used by Josh to combine the A and B outputs to feed his HP amp.

==> Doesn't look like a professional design to me. OK, it is over 10 years old and likely was never intended to be a top-tier device for line level ABXing, but still...

So far my conclusions would be
a) The ABX box itself likely has issues.
b) The setup in general maybe had issues (like polarity, etc).
c) The procedure certainly had issues (no controls used, etc).
And because no recordings were made, we'll never know.

Very good eye!

From the specs for that part: "Low resolution DAC replacement."
View attachment 547520

Yes, we are running the output of high performance DAC through a low performance one to adjust its level!

Author says that the effect is constant on both inputs. Not so unless the level of attenuation is the same. Which was not the case or else it would not be needed!

We are talking this level of non-linearity here:

View attachment 547519

This is just not the intended use, i.e. test very high high performance sources. It is made to compare amplifiers and such.

I'm catching up and saw this. And, oh no...
EDIT : no no, all the below is wrong, I should have read KSTR before. I need a break
And there's that one too:

So, with a 2Vrms input and an INL of -0.1LSB to -0.3LSB, even if the full scale output is level matched at 0.1dB, the above means a potential variance of up to 0.2dB at -20dBFS, that would be a quieter moments in music. That with high distortion and crosstalk too, also virtually increasing the noise floor of one of the two devices under test if they required level matching.

No need to investigate further, that alone means the real device under test was the box, but we suspected it from the beginning.

Select a fade to noise moment, push the volume and get 100% result.

All the DACs in use acted as "perfect generators" to feed this low resolution device.

Bottom line: It was a not so interesting testing of the AD5262 digital potentiometer. Reading the datasheet would have been enough to guess the final results in the context. Like Erin wrote, the box can be useful to play with 2 amplifiers of very different output impedances to hear something. Beyond that...