Denafrips DACs might not actually be NOS?

[GoldenOne]

So I've got an Ares 2 in for review from a friend at the moment, and whilst testing it I've noticed that the NOS mode is...well...not actually NOS
(All images are taken from 768khz recordings of balanced output of Ares 2 in NOS mode)

Impulse response looks like this:

 


The rise/fall rate changes depending on sample rate, indicating it is indeed DSP/OS and not a slew rate limitation

44.1khz:

 

192khz:


And lastly, putting a 1khz sine @ 44.1khz sample rate yields a result that doesn't exactly look like NOS to me:



If anyone else here has a denafrips DAC and an ADC that can do 192khz or higher, could you check if yours is the same? Lmk if you need test files

[GoldenOne]

1 minute ago, The Computer Audiophile said:

Hi Guys, I heard back from Alvin, who spoke with Denafrips. He says, according to Denafrips, the NOS is indeed non-oversampling. He couldn't provide more information because Denafrips doesn't want to reveal anything proprietary with respect to how its DACs operate. 

 

I'm not sure where this leaves us, but it's all I have right now. 

I guess maybe they could call it 'non oversampling' on a technicality in this situation given as the FPGA is just making things move in a linear fashion to the next sample.
Technically not adding extra PCM samples perhaps but still functionally oversampling/interpolating.

We know it's definitely not a slew rate or analog low-pass filter limit because the square wave transfer speed changes depending on sample rate.


The linear interpolation behaviour can be seen quite clearly on a 15khz sine:




For those who are unfamiliar this image shows what is happening in the denafrips impulse response vs what would happen in true NOS/Zero-Order-Hold behaviour (blue line):
 

- Sample 1 is received by the DAC, nothing happens and it holds at 0 until the next sample.
- Sample 2 is received by the DAC, it immediately moves up to the value of sample 2 and holds until the next sample
- Sample 3 is received by the DAC, it immediately moves down to the value of sample 3
- Sample 4 is received, same value as sample 3 so voltage stays where it is.
 

BUT, that isn't happening on the denafrips.
Instead, immediately after sample 1 arrives, it begins moving up toward sample 2. The only way this could happen is if additional samples had been interpolated in-between samples 1 and 2, ie: it was oversampling.
It is using a filter that introduces no ringing, but it is nonetheless still oversampling.


In reality a NOS impulse will look something more like this because most R2R dacs will have some form of analog reconstruction:

[GoldenOne]

16 minutes ago, The Computer Audiophile said:

 

This is interesting. 

 

1. I wish we had more information from Denafrips. 

2. I wonder what the pros and cons are of doing it the way Denafrips appears to be doing it.

3. If it's not a traditional NOS design, I wonder why and if this is better or worse than a traditional NOS design. 

 

All interesting stuff!

 

1) Yep definitely. I understand they won't want to disclose exactly how their DACs are working. Given they are indeed exceeding the accuracy of their resistors it shows that certainly some kind of compensation is both at play and working and they wouldn't want to spill their secret sauce, but it would be good to get this cleared up given as it's one of the selling points of the product

2) I guess the pro would be that it might allow for more flexible correction techniques or subsample correction with the FPGA where it might not be possible with NOS?
The con being that because it is mostly randomly (but of course in a signal dependent manner given as it depends where the samples are as to where the interconnecting interpolated line moves to/from) creating additional tones lasting 1 sample duration, this COULD maybe have an effect on noise floor with musical content?
With true NOS the transition is always just going to be vertical/near-vertical and so will remain constant. But then on the denafrips interpolation it would be random according to the signal being played.
Measurements seem fine, but given as FFT relies on content being present for long durations (hence why you can use it to look below the noise floor), the effects of what denafrips is doing here might not actually be particularly visible on an FFT.

To be honest I think the best person to ask might be @Miska. What in your opinion would be the major drawbacks of using fully linear interpolation like what seems to be happening here?
And are there any particular benefits?

[GoldenOne]

41 minutes ago, The Computer Audiophile said:

 

Agree, hopefully Jussi has time to add an opinion here. 

 

By the way, I like your approach to investigating this one. Curiosity is always best when we don't have all the information, rather than finger pointing and assuming something sinister. Bravo. 

Absolutely. 
I just do this cause I'm curious and find it interesting and am not at all pointing fingers. Just wanting to know more :)

To be clear it DOES look like what they're saying on the tin isn't strictly true, at least not in the traditional sense. And I think that that might not be a good thing.
But it's certainly not as clear cut as in other situations where it's just a binary "is it lossless" question.

There could be more at play here and honestly I know enough to know when I don't know enough. This is one of those times and I'll leave it to people with better understanding than myself to suggest the answers. I just wanted to start the conversation. It's an interesting middleground area between software/digital stuff in terms of oversampling/DSP, but also hardware in terms of controlling and compensating a resistor ladder.

Miska is most likely the best person to ask about the OS/Software side of things. But I don't know if we'll ever know exactly what's going on here in terms of the hardware side for sure.
I just found it interesting and thought others might too.

Also, seems like in the @John_Atkinsonstereophile review they presented some interesting behaviour when he tried to measure linearity. It was oddly deterministic. So that could perhaps further suggest something going on in terms of DSP.

[GoldenOne]

2 hours ago, idiot_savant said:

Linear Interpolation gives a squared sinc frequency response - if you look at John Atkinson's measurements, you can see it gives you -1.5dB @ 10kHz, and -6dB (ish) @ 20.0k. NOS gives you a sinc frequency response

Doing the sums:

PI/4 ( so 11.025k @ 44.1k sample rate ) SIN(PI/4) / (PI/4) gives us -0.91dB droop  for a NOS

PI/2 ( so 22.05k @ 44.1k sample rate ) SIN(PI/2) / (PI/2) gives us -3.9dB droop  for a NOS

 

PI/4 ( so 11.025k @ 44.1k sample rate ) (SIN(PI/4) / (PI/4))^2 gives us -1.82dB droop  for a linear interpolate

PI/2 ( so 22.05k @ 44.1k sample rate ) (SIN(PI/2) / (PI/2))^2 gives us -7.84dB droop  for a for a linear interpolate

 

So from everything we've seen here, it's a linear interpolate.

 

Your friendly neighbourhood idiot

Thanks for this.

Just checked and yep the ares 2 rolls off quite a bit more than the may

 

[GoldenOne]

I just want to make clear, that I made the original post because I thought it was interesting, and in the hopes that either denafrips or someone else would be able to provide further information or suggestions for testing. Which is what @idiot_savantdid (thank you for that).

This is not and was not intended to be a "Denafrips bad" post. If they are lying about the DAC being NOS that is indeed concerning. But given as the DAC sounds fantastic in OS mode and most people seem to prefer it in that mode anyway, the content of the OP is entirely separate from whether or not this is a good dac. (Which it is)

You can have a constructive discussion or criticism of a product which you like. There are plenty of products I love which I have issues with either in terms of design or business practices, and plenty of products I dislike which also have attributes I am fond of or appreciate.

There is no need to lump things into a "100% good" or "100% terrible" box

[GoldenOne]

21 minutes ago, numlog said:

To me it looks like linear interpolation aswell, a very primitive form of OS but OS nonetheless. It would help to reduce the amount of ultrasonic content on the output compared to NOS but comes at the cost of high-end roll off. A frequency response graph could confirm that. Probably done to improve measurements.

 

 

 

Post on that is here:

 

[GoldenOne]

1 minute ago, The Computer Audiophile said:

I have no clue why anyone would use anything in NOS mode without using external over sampling. 
 

I send 1536 kHz to the Terminator in NOS mode, using HQPlayer filtering. 

I think it depends on the device. 

 

I actually really like NOS on the may (which has some analog reconstruction). But on the ares 2 I don't at all (though I don't think it is actually nos as discussed). 

 

It'll be interesting to try nos on the phasure NOS1a if I can actually get it working. At the moment it's an H shaped paperweight

[GoldenOne]

1 minute ago, idiot_savant said:

Well, the paper @etanequoted pretty much advocates NOS at 44.1k because it’s too noisy otherwise. I think it’s a thing that people like that I don’t understand.  
 

As I’ve said before, linear interpolates can be a useful tool ( so eg I think Chord use one in the final interpolate ) but not at 44.1k. 
 

IMHO the problem here is that you have a DAC that ( as it stands ) you can set it up to be really quite poor in frequency response terms, then as you increase sample rates via eg HQPlayer it suddenly sounds much better, when it could have been perfectly fine already…

 

your friendly neighbourhood idiot

I think technically though if upsampling to 705.6/768mhz then the dac filter itself makes very little difference regardless of design (so long as its not doing something completely crazy). Probably more just a processing power vs benefit question when going from the 768khz source to whatever is actually being fed to the dac

[GoldenOne]

15 minutes ago, idiot_savant said:

@GoldenOne you are, of course correct, in that filters should have less effect the higher in sample rate you go, so that brings us back to what is the point of a NOS mode at lower sample rates?

do you happen to know anything about the Holo analogue reconstruction filter? I can’t see it doing much good unless matched with an OS filter?

 

your friendly neighbourhood idiot 

Not too much no. Though it's not intended to be a replacement for actual oversampling.

 

The may has internal oversampling if desired (and can also upsample to dsd and play back using its dedicated 1-bit converter) and also has 1.536mhz/dsd1024 input support so that you can do exceptionally high rate OS with software like hqplayer. 

 

Its got all the options so the user can use it however they choose. 

 

The NOS mode is indeed NOS and seems to just have a gentle analog filter to get rid of ridiculously high stuff (>50khz etc). Though I believe @Miskahas mentioned that the holo dacs may be doing something more advanced than a basic low pass filter. (could be wrong there but I remember reading a mention of that) 

[GoldenOne]

I think this topic may have gotten sidetracked slightly :P

Firstly, what happens BEFORE the dac does not inherently change what the DAC itself is. The Phasure NOS1 is a true filterless NOS DAC, and the Holo May is a true NOS dac with some analog reconstruction.

Both of these are genuine NOS, and both of these can be used with oversampling software such as XXHighEnd or HQPlayer. In fact the NOS1 it's specifically recommended to use XXHE.
But using software oversampling does not change the fact that the DAC itself is indeed NOS. So this discussion about software is separate.

There are lots of different things that can alter the output. PERSONALLY, I'd consider NOS to be anything that in terms of the actual input to the DAC, adheres to non-oversampling/sample-and-hold. Meaning it moves to the value of the sample, and then holds there until the next sample. Anything else is not NOS as there is additional interpolated data being created and used.

Though I would say that subsample correction like what Audio GD mentions doing is probably still fine. Correcting for ladder inaccuracies whilst still leaving the actual sample data unaltered is probably fine. Though I suppose up for debate.

And analog reconstruction is...well...analog, so there's no question that it's unrelated to digital sampling.

Impulse from Phasure NOS1, complete true NOS with no filter:


Impulse from Holo May, true NOS with analog reconstruction:


Impulse from Denafrips, OS, Linear interpolation:

🎵One of these things is not like the others🎵

(And again, to be very clear, this isn't at all me 'bashing' denafrips. I like the ares 2 a lot. But I do believe their description of "NOS" is untrue. It is linear interpolation as all evidence shown demonstrates. I'm also NOT implying that either NOS vs OS or the variations of either is "best". I like both NOS and OS for different situations and think both have merits.)

[GoldenOne]

2 minutes ago, idiot_savant said:

@GoldenOne do you still have the Ares 2 there?

 

your friendly neighbourhood idiot

Yep I do

[GoldenOne]

1 minute ago, idiot_savant said:

@GoldenOne and do you have HQPlayer?

 

your friendly neighbourhood idiot

I do yes

[GoldenOne]

3 minutes ago, idiot_savant said:

Excellent! I'm assuming you can do FFTs and generate test tones?

Lets do an experiment:

Let us generate a 44.1kHz 24 bit -60dBFS 1kHz sine wave

 

In all cases, NOS mode on the DAC.

 

1)play that using HQPlayer no oversampling, grab FFT @ 768kHz

2)play that using HQPlayer, 1.411MHz, grab FFT @ 768kHz

3)play that using HQPlayer, 1.411MHz, noise shaped ( I don't know the options - 5th order? ) to 16 bits, grab FFT @ 768kHz

4)play that using HQPlayer, 705.6kHz, noise shaped to 16 bits ( same settings ), grab FFT @ 768kHz

 

I'm honestly guessing at the 16 bit number here, it would be helpful if @The Computer Audiophile shared his settings?

 

If you could, that would be most excellent,

 

your friendly neighbourhood idiot

 

 

 

 

I'll do this shortly. Though keep in mind my ADC only goes up to 768khz so when running at 1.536mhz on the DAC I may not actually be able to detect all that much, but will see!
This is the kind of area where an analyzer would come in very handy. I'm working towards getting an APx555 at the moment, all patreon income is going towards that. That's got a 1Mhz bandwidth so would be great here

Additionally though, I'd say that when feeding a DAC 1.536mhz it almost doesn't matter what filter it's running as the information is so high bandwidth anyway. The discussion/OP was about what happens to 44.1khz content. HQP brings a benefit to D/S dacs with oversampling so I'm sure the same would happen here.

[GoldenOne]

So I put the denafrips IR through the APx555 today to look into things a bit further and found a couple interesting tidbits.

When putting an impulse response through it, viewed with a 2.496mhz sample rate, we see this:
These 'steps' appear to be 32x the input 44.1khz sample rate.

The second interesting thing is that the impulse response derived from the transfer function looks like a normal sinc impulse response. Whereas deriving the IR on either the holo may or the ADI-2 with the 'NOS' filter there, shows something very close to the true IR.

So there is definitely some seemingly adaptive DSP or something at work here.

 

[GoldenOne]

1 minute ago, idiot_savant said:

Interesting - this is almost certainly *not* processing delay as the only way you'd get this much ( 40ms ) is thousands of taps, which isn't backed up by the stereophile measurements.

I'm sure @GoldenOne would love an excuse to measure with his shiny AP?

If I had to put money on it, I'd say there's some kind of FIFO scheme in there. Hmmm, this might be interesting - I reckon the AP can measure group delay, and then graph that over time ( like minutes ) - do you fancy this?

 

 

 

your friendly neighbourhood idiot

 

The AP has a DUT delay test so I'll run that in a sec (currently rendering a video so might need to wait a bit).

Also the AP isn't mine it's just being loaned to me temporarily. Still a while off getting one permanently but working towards it.

[GoldenOne]

2 minutes ago, idiot_savant said:

If possible, using the same input as Darko ( i.e. Toslink? ) - although I think AES would do.

 

Should think the loan unit has got you salivating for your own!

 

Although didn't JA at Stereophile have his AP loaned from them for years and years?

 

 

 

your friendly neighbourhood idiot

It most definitely has!
Using it it's immediately clear why AP is the industry standard. It's almost nothing to do with the performance, it's the combination of hardware and software flexibility and how even custom, complex tests can be set up and run in a matter of seconds. It's immensely powerful.

I'm not sure what JA's arrangement was. But a large publication like stereophile is likely a fairly different sort of deal than a moderately sized youtube channel.
At the moment my intention is to purchase one as soon as funds are available. I'm looking at if it would be possible to get one on a hire-purchase scheme as that'd make things much easier but being a new business and me personally being quite young it's unlikely.

This unit has been kindly loaned to me for a couple weeks by AP to check that it will do everything I need it to do (and to fully sell me on it of course in case I had any doubt. With something as impressive as this a hands-on test really does let the device sell itself).

I'll ask them if there is any possibility of being able to keep ahold of a unit, maybe even sort out some sort of sponsorship arrangement. But honestly the tricky thing is that AP really doesn't need it. They're the industry standard for a reason and anyone who might need an AP already knows them. I doubt they need much advertising and given the cost of the device I doubt they'd see the value in it. Stereophile MAYBE, ASR paid RRP, and so I can't imagine I'd be much different.

In any case, I'm determined to get one. I don't know exactly when it'll happen, but as soon as I'm able to do so I will.

[GoldenOne]

1 hour ago, idiot_savant said:

Interesting - this is almost certainly *not* processing delay as the only way you'd get this much ( 40ms ) is thousands of taps, which isn't backed up by the stereophile measurements.

I'm sure @GoldenOne would love an excuse to measure with his shiny AP?

If I had to put money on it, I'd say there's some kind of FIFO scheme in there. Hmmm, this might be interesting - I reckon the AP can measure group delay, and then graph that over time ( like minutes ) - do you fancy this?

 

 

 

your friendly neighbourhood idiot

 

Ares 2 OS and NOS have near identical delay. The values shown are consistent and did not change from run to run.



ifi iDSD diablo for a quick comparison:




So yes, it seems that the Ares 2 does indeed have a notable delay. Nothing 'huge', but definitely enough to throw off a movie audio sync.
The above were taken using SPDIF given as that is what Darko used.
USB may be slightly higher if there is an additional buffer there. (Can't test that with the AP)

[GoldenOne]

1 hour ago, idiot_savant said:

@GoldenOne- can you maybe leave it running for a bit? We seems to have a 20ms or so discrepancy - I’m interested if the delay changes over time?

 

 

 

Your friendly neighbourhood idiot

Well it's a good thing you said this!
I just connected it again via coax this time. I was getting ~25ms at first, but I just checked again 2 mins later aaaand:



Thank you for suggesting that!
I'll leave it connected for a bit and try again shortly.

Why is it do you think that it gets longer after a while?

[GoldenOne]

I don't have the option to do a DUT delay/time test setup so am just running it manually every few mins, but it is steadily getting longer.

Currently at 129.8.

I'm gonna go grab some food then will see what it's at once I'm back!

[GoldenOne]

4 hours ago, numlog said:

Might be interesting to see if 32x steps are visible in OS mode with different input data rates, that would probably indicate if there is any difference between NOS and OS mode when fed 1.4Mhz data, and if it is capped at 1.4Mhz.

Stepping behaviour is indeed visible in OS mode.



 

[GoldenOne]

Ok so the delay seems to be....somewhat random, but also sloowly gets longer over time (not sure if there is a cap as such but we will see!)

First go: 25ms, moves to 129ms after a while
Unplug, move to coax 2: 85ms
Unplug, move to coax 1: 81ms
Swap back to optical: 163ms
Swap back to coax: 140ms

Leaving it on the coax now and will do for a short while.
140.5ms currently

[GoldenOne]

23 minutes ago, opus101 said:

I'm wondering if the variable delay has something to do with their invincible PLL? If PLL is off does the delay revert to a stable number?

 

The Mola-mola DAC I understand also has a somewhat variable delay but not one which changes over time. Just each time it locks the delay is indeterminate. This I gleaned from posts by one early adopter over at WBF.

It shouldn't do.
PLL shouldn't add any latency. A buffer or DSP would though.

Holo may latency for comparison with PLL on (blurred bottom result cause only one channel was connected):
 

[GoldenOne]

4 minutes ago, idiot_savant said:

My concern here is that latency growing over time is actually a sign of a lack of a PLL… be interesting to see if there is an upper limit. Oh, one other thing to check - with a 1kHz tone, does the AP measure it as 1kHz?

 

 

 

your friendly neighbourhood idiot 

Yep it is measuring as 1.0000khz so no issue there.

Short video of the delay thing here: https://streamable.com/vnz7j8

When the device is initially connected the delay is seemingly random. I've seen as low as 4ms and as high as 190ms.
I've got it at 155.6ms right now and am just going to leave it plugged in for the rest of the day to see what happens

 

[GoldenOne]

1 minute ago, idiot_savant said:

OK, so I think we could have 2 issues here, which we should try to separate:

 

Issue 1 : seemingly quite a big variation in latency between inputs changing or going away. 

Issue 2 : latency seeming to increase over time. 
 

issue 1 is not great and will irritate people trying to watch video, but apart from that, nothing to see

 

issue 2 is more problematic in terms of what’s going on - *how* is the latency increasing? At what rate? Is  there a maximum and minimum?

 

so, if you could just leave it running the continuous sweep doohickey and look at latency every say 15 minutes, we should be able to work out the rate latency is increasing, and if there is a maximum

 

 

your friendly neighbourhood idiot 

So with #2, it seems that it's not always increasing.
In fact this run I started at about 156ms, and it has slowly decreased to 153ms over a few hrs.

So it seems that it changes one way or the other but isn't always increasing.