Review and Measurements of Benchmark DAC3

[amirm]

I'm not sure that it's sorted. Let's say one channel is off by a constant (say) 3uV. So 2.0 v in one channel corresponds to 1.999997 in the other channel. At that 2V, the difference is 1.3 exp -5 dB, which just won't be indicated on that graph. At 1mV, the "off" channel will be 0.997 mV, which is 0.02dB, which will show up. At 10uV vs 7 uv, it's 3.1 dB... you get the point. Both channels could be tracking perfectly, but the log nature of the plotting will make it look as if the offset is a nonlinearity.

Can you eliminate that possibility (or punch a hole in my logic)?

Any constant offset is taken out of the measurements. Before anything starts, a measurement is made at -20 dBFS. Then all computed measurements are relative to that. So if the offset is constant, it will also be in -20 dBFS so it will not be material.

Each channel is also independent so it doesn't matter if the other channel has offset or not.

 

[Jimster480]

@Jimster480, maybe I've missed something but why do you keep calling the Benchmark DAC3 "UDAC3"?

That's a good question... I think I am just doing too many things at the same time and confused the product name tbh.

 

[mindbomb]

@amirm

One thing I was wondering was about the toslink performance on the rme adi 2 pro. I often see on AKM dacs, they use the ak4113 for voltage reasons, and the noise and jitter performance drops compared to usb.

 

[amirm]

@amirm

One thing I was wondering was about the toslink performance on the rme adi 2 pro. I often see on AKM dacs, they use the ak4113 for voltage reasons, and the noise and jitter performance drops compared to usb.

I will try to remember to test that next time I am at the bench. Remind me if I forget.

 

[Blumlein 88]

Any constant offset is taken out of the measurements. Before anything starts, a measurement is made at -20 dBFS. Then all computed measurements are relative to that. So if the offset is constant, it will also be in -20 dBFS so it will not be material.

Each channel is also independent so it doesn't matter if the other channel has offset or not.

I suppose this is similar to what I do. I have a tone at the beginning at -60.0 dbFS. I'll turn up the input gain on the ADC to get well above its noise floor. Then I take the recorded file and reduce it until the tone is -60.0 dbFS again. So any offset I should have adjusted out that way.

 

[March Audio]

I've the reverse opinion. I agree with AP. If we're doing linearity then do that. We know the noise floor will interfere even if the chip puts out a signal amongst the noise of the proper level.

I'm with you on this one. I want to exclude noise to understand what the underlying linearity is. These are two separate issues. Noise is more usefully observed in the time domain. Those - 90dB sines clearly show which dacs have good noise performance, but of course it can be measured in other ways.

The problem with trying to include noise in the linearity plot is that there will always be some integration of this noise dependant on measurement parameters. If looking at a Fft value, the higher the resolution the more the noise is smoothed. If measuring a voltage directly it will still be integrated.

Too many variables, including of course the APs own noise. However by the method I use, Highly averaged high resolution Fft, you obtain consistent levels of the test tone excluding, as far as possible, the noise variable.

 

[SIY]

Any constant offset is taken out of the measurements. Before anything starts, a measurement is made at -20 dBFS. Then all computed measurements are relative to that. So if the offset is constant, it will also be in -20 dBFS so it will not be material.

For my theoretical 3 uV offset, the difference at -20 dB would be 0.0001dB. Would the measurement at that level be able to account for that? Are there enough significant figures?

 

[FrantzM]

For my theoretical 3 uV offset, the difference at -20 dB would be 0.0001dB. Would the measurement at that level be able to account for that? Are there enough significant figures?

We're really splitting hair here...

 

[SIY]

There's a French expression regarding performance of unnatural acts on flies, but I'll decline to use it.

But you get my point- what's a hair at high level, if constant, becomes a thick rope when you go down 5 orders of magnitude (-20 to -120 dB).

 

[DonH56]

It reads like the AP is doing noise averaging (to reduce the noise) during the linearity sweep (?)

 

[ManglerFi]

That's a good question... I think I am just doing too many things at the same time and confused the product name tbh.

Ah, I see. No worries, I was just wondering if there was a reason for it. Thanks.

 

[Simon740]

Hello,
I’m looking a good dac with good volume control for my active speakers. Everyone use this dac with active speakers?

Regards
Simon

 

[amirm]

It reads like the AP is doing noise averaging (to reduce the noise) during the linearity sweep (?)

APx555 out of box has a simple linearity measurement with no filtering but only goes as deep as -65 dB. Here is the graph by default:

As you see it goes from -60 dB to +5 for a total of 65 dB from zero reference. In this use, no filtering or noise reduction is necessary as they are well below what the template was designed for.

The moment you dial this way down to -120 or heaven forbid, -140 dB, you are on your own to figure out how to create correct measurements.

APx555 provides a number of canned filters with different responses that can be used including a full custom one which is what I have created. Either way, the measurements are in the domain of what the operator does than what AP provided.

For my older AP SY2522 analyzer, Audio Precision provided the full solution in the form of cascaded digital and analog analyzers with bandpass filtering. That was nice because we could all use and follow the same script and there was nothing left to interpretation/setup.

The scheme for getting rid of offset was also different in that the older AP would try to use a least error method to use as a reference whereas the new APx555 uses that -20 db (or whatever you set) as reference.

Finally, for ALL measurements in AP, behind the scenes there is a settling algorithm that attempts to make sense out of run to run variations. The older analyzer had the same but due to differences in underlying measurements, I find that the APx555 gives up on finding the optimal value when the measurement values are small (whether it is THD, IMD, or simple amplitude measurements). We are dealing with very small values in our testing of DACs which apparently is not a common scenario. Fortunately the parameters can and should be adjusted for values to converge.

This is the reason I say that the claim that APx555 is better than 2522 and hence anything it shows is more accurate then 2522 just doesn't hold. If you want to compare those sets of measurements you need to set up the APx555 so that it closely emulates what the 2522 did. Otherwise you have two different measurements.

 

[amirm]

Hello,
I’m looking a good dac with good volume control for my active speakers. Everyone use this dac with active speakers?

Regards
Simon

While I did not try it, of course you can do that. It comes with a remote so you can adjust the volume easily.

 

[amirm]

I'm with you on this one. I want to exclude noise to understand what the underlying linearity is. These are two separate issues. Noise is more usefully observed in the time domain. Those - 90dB sines clearly show which dacs have good noise performance, but of course it can be measured in other ways.

The problem with trying to include noise in the linearity plot is that there will always be some integration of this noise dependant on measurement parameters. If looking at a Fft value, the higher the resolution the more the noise is smoothed. If measuring a voltage directly it will still be integrated.

I want to emphasize again that substantial amount of noise reduction exists in my measurements. Levels are attenuated by a whopping 50 dB on either side of the frequency of the generator. Without it, you just measure the noise, not the signal.

We want to use time domain analysis here because AP runs in automated mode there. So we can make the 60+ measurements in linearity mode or whatever we like. Trying to do that in FFT manually gets very tedious. Hence the reason we are interested in proper filtering for measurements in time domain.

The other but related issue is that all filtering in APx555 are digital. As such, you are dealing with ringing and other non-ideal response than what one imagines. But not using too steep of a filter, I have been able to minimize these.

 

[etc6849]

Thanks for these measurements Amir! Please keep up the great work.

I am surprised by the unbalanced performance of the DAC3. Good thing I plan to use the balanced outputs on my DAC3 DX when my HPA4 arrives the week after next!

I think I made it through all 10 pages, but just in case, were the channel differences noted on the balanced outputs alleviated when the calibrated output was used (as fed from USB and/or SPDIF)?

 

[amirm]

Thanks for these measurements Amir! Please keep up the great work.

I am surprised by the unbalanced performance of the DAC3. Good thing I plan to use the balanced outputs on my DAC3 DX when my HPA4 arrives the week after next!

I think I made it through all 10 pages, but just in case, were the channel differences noted on the balanced outputs alleviated when the calibrated output was used (as fed from USB and/or SPDIF)?

My pleasure. I only tested the linearity with and without calibration mode. Which level were you interested in anyway?

 

[etc6849]

I am planning to use the DAC3 DX as a USB DAC with calibration mode enabled for the XLR outputs (to eventually feed the Benchmark HPA4).

I found your unbalanced linearity measurements showing calibrated mode on versus off; I was just wondering if there were similar comparison for the balanced output.

My pleasure. I only tested the linearity with and without calibration mode. Which level were you interested in anyway?

 

[amirm]

No, I did not re-test that since balanced was already excellent. I had forced attenuation to balanced level anyway using the jumper.

 

[svart-hvitt]

@amirm , it just struck my mind: It would be interesting if you got hold of the Benchmark Dac1 and Dac2 to make comparisons; how does the «same» design measure over time. This would cast light on what’s been the development in the dac area over the past decade.

Benchmark is held in such high esteem among «objectivists» so an mk1, mk2 and mk3 comparison would be of high academic interest.