Superclocks

[Miska]

On 8/6/2018 at 5:25 AM, AudioBang said:

I'm posting my experience upgrading the Crystek CCHD 575 100MHz oscillator [$25] in the Vega DAC to the Pulsar [$425] merely for perspective for anyone interested. Conclusion: In this application, reducing the power supply noise made a significantly bigger improvement over TXCO to OCXO upgrade at 10X the cost.

 

The actual measured phase plot provided with the Pulsar showed a 10dB improvement at 10Hz and 100Hz over the Crystek CCHD-575 generic datasheet published on their website.

 

In many cases I've seen, jitter performance of DAC output doesn't change much by changing the clock part. Because the problems are more in the PCB layout design and other parts rather than the clock module itself. There's not much joy about fancy clock if it is spoiled before it reaches the DAC chip pin... And even if it reaches the DAC chip pin unspoiled, the it can still be spoiled by noise coming to any other pin of the DAC chip.

 

So I'd recommend to stop looking at datasheet plots and start looking at jitter measurement output of the entire DAC device instead. Because the two don't often correlate much...

 

[Miska]

7 hours ago, barrows said:

Jussi, the interesting thing is, swapping to a better specified clock has always worked for me in terms of subjective listening, with a clear improvement, but the J-test spectrum looks the same.  I suspect the J-test is insufficient to see what is really happening, after all the J-test was developed basically to to test (the rather poor) SPDIF interface performance, we need a better measurement.

My experience is with changing clocks with the same implementation as far as clock distribution to the DAC chip goes.

 

I think J-test is very good, since it tells if there's pollution on the clock. A DAC with fancy clock may perform poorer than another DAC that has regular clock. I prefer to pick the DAC that performs better, not the one that has fancier part inside.

 

Problem in changing clocks is that it is hard to place new clock properly in regards to PCB layout. For example if a DAC is originally using the NDK clock part and then you would like to swap to Crystek, no way it is going to fit there. If you put any wires there, then you are certainly big time spoiling the PCB layout and likely adding a lot of jitter. So at least the performance needs to be always verified by measuring, to check that it doesn't actually get worse.

 

Can you post some J-test24 measurements what is the baseline performance of the DAC where you are swapping the clocks and it makes difference sonically? I would like to see how close they end up realizing the clock performance in practice.

 

[Miska]

2 hours ago, barrows said:

But with the J-test, the spectrum generally appears perfect, even with a relatively "standard" XO...  

 

I'd be happy to see first DAC with perfect J-test figure! Please post!

 

So far, almost all problems I've seen are related to something else than the clock module itself. 99% of the problems are related to design of the surrounding circuits or the PCB layout (or just the PCB materials). So I'd rather spend the efforts improving where the problems are (PCB layout design and surrounding circuit design), rather than messing around with a component that has very minimal effect on the performance compared to aforementioned aspects.

 

Higher the clock frequency is, worse it gets anyway. So better to use 22.5792/24.576 MHz clocks with DSD512 without clock dividers (there are such DACs)! If you go for clocks near 50 MHz you already suffer about 10 dB in phase noise. Let alone clocks near 100 MHz needed for ESS Sabre. You could even optimize and use 11~12 MHz clock and DSD256 and still have flat noise floor up to over 100 kHz.

 

2 hours ago, barrows said:

I wish I had access to an analyzer which can measure the phase noise, it would be very interesting to measure XO phase noise at the input pin of the DAC (chip), but the expense of that tool is not justified at this point.

 

No reason to bother with such, because it would be irrelevant. Only relevant thing is what comes out of the DAC's analog output. Even if you have perfect clock in the clock input pin, but if you have USB packet ticking and a bit of extra thermal noise leaking to the reference voltage pin the entire result is still spoiled.

 

2 hours ago, barrows said:

One thing I saw recently, on a not to be named audio review site, was a very expensive Ethernet switch a company was selling, which included an "upgraded" OCXO.

 

I have to admit I never understood why mess with such in first place. If you use for example HQPlayer with NAA, or just a generic UPnP Renderer, the network traffic doesn't carry any clock information what so ever. When you listen to Tidal streams the jitter over internet is already in several milliseconds range, and still it doesn't matter at all. NAA can handle jitter up to about +- 500 milliseconds.

 

2 hours ago, barrows said:

I can have the J-test measurements you suggest made with different XOs, but the AP is in Florida and I am in CO so it takes some time.  I may get the AP sent here if/when I get an even better clock for comparisons (Pulsar).

 

Did you already mod the clock in AP too?

 

[Miska]

19 minutes ago, barrows said:

Well, i would say we disagree on this.  And checking for noise on the DAC's voltage reference is pretty easy, so easy to rule out as a source of problems.  USB packet noise is 8 kHz, I do not see it, there is pretty good isolation from that noise in this DAC with dedicated ADM 715x regs for each DAC Vin, and isolation between USB receiver and DAC I2S input.

BTW, I am not modding commercial DACs with alternate XOs, I am building DIY DACs and evaluating different XOs with the same implementation (for clock distribution).

 

Checking the actual DAC analog output is easiest and the only relevant thing. It tells the actual facts. Because there can be also poor/buggy chip designs...

 

I cannot comment on anybody's DIY DAC, only on commercial DACs. And the weakest point is usually not the clock module. And I have a lot of doubts about modifying commercial DACs. Replacing a clock in tightly packed SMD board of a commercial DAC is not straightforward if it is not exact footprint match. And even if it is, it could have some characteristic differences not taken into account in the original design. Blindly going and replacing a clock module to a "better one" on a commercial DAC and hoping it'll improve things is likely not going really do that.

 

And there are other funny things, like for example ESS' wandering noise bumps around the noise floor likely coming from the ASRC. These are actually hard to catch because they are time-varying and very low level.

 

19 minutes ago, barrows said:

As for "perfect" J-test results with regards to jitter spectrum, OK, perfect is in quotes!  But take a look at the Stereophile measurements of commercial DACs, these days a lot of them have virtually perfect results.

 

I always read though all of them, but somehow I haven't yet spotted any perfect result. I'm not interested in 16-bit J-test results because that hides a lot. 24-bit is the relevant one. Of course I don't read JA's comments, just what I see myself. I also was checking through all the HiFi-News (UK magazine) results when PM has been publishing the results. And of course I measure all the gear I happen to get my hands on.

 

[Miska]

54 minutes ago, barrows said:

Do you not hear differences in various HQPlayer filters which produce the "same" analog output (in the audible bandwidth) in standard measurements?  Just curious. 

 

I have no problem analyzing/measuring differences between HQPlayer filters and modulators...  It is not right/wrong, just different.

 

It is not something that would elude measurements but be audible subjectively. Sometimes the differences are easier to analyze than to hear.

 

My design flow is such that I do objective evaluation first and only once it has passed, I start to listen. I don't even want to listen to something that doesn't look as expected on analysis.

[Miska]

8 hours ago, barrows said:

You have measured differences within the range which would be audible?  That is <20kHz and above -110 dB?  Really?  I mean, I get that an early rolloff filter could be considered a measurable/audible difference (although not theoretically for my hearing)-but what about minimum phase vs. linear, for example...

 

I don't want to specify what figures are "audible", I don't have clear answer for that. Not for clock jitter/phase noise either. I just know what is measurable or detectable in digital domain analysis. For example filter transition for RedBook begins somewhere below 22.05 kHz when it begins to roll off from 0 dBFS. This is certainly measurable in frequency response. Minimum phase vs linear phase is certainly measurable in the phase response too.

 

[Miska]

8 hours ago, mansr said:

Making a filter/modulator that causes such differences is trivial. Then again, so is making one that doesn't.

 

I don't have problems analyzing differences of the stuff in SoX. Your different modulator each produce different results. If two things produce same output values, then they are usually the same.

 

I just don't put any limits like "20 kHz" or "-110 dB" because I think such limits are artificial. What I look at, only limit is computational accuracy or limits of measurement. But usually the filter roll-off point and roll-off curve steepness and shape ("shape of the knee") is certainly measurable.

 

[Miska]

3 hours ago, barrows said:

of course I understand that, but is it audible? certainly a measurement showing the rolloff above even at 19 kHz is not audible, But still people hear the difference in filters-I would conclude that something else about the filter's response is what we hear, but we are not measuring that thing.  This is the point which I was trying to make.  Same thing with linear phase vs. minimum phase, the phase shift at a very high frequency which we can measure, is also not considered audible by humans (unlike low frequency phase shifts), but still people report differences in sonics of these two filter types (overly simplified here, as we can have intermediate phase filters as well).  We can measure the pre-ring/post ring distribution with an impulse response test, but is that ringing audible?  Very questionable that it is.

 

My take is that it is audible. You know I do endless cycles of adjusting the mathematics, analyzing the result and then listening. I have my own idea which things are audible in which way, just based on my own experience.

 

And I have many cases which I would count as "blind listening", because I get a lot of feedback, and in many cases people simply cannot know what is behind. That has shaped my view also what is audible.

 

I think the scientifically defined limits of human hearing are very average and the research methodology there would need more improvement than measurements.

 

3 hours ago, barrows said:

I am just trying to illustrate that the measurements we currently use are inadequate to describe all aspects of the sound of (in this case) a DAC, and that for similar reasons, perhaps the higher performing clock, not showing any difference in an analog domain J-test measurement, is still affecting some aspect of the sound.

 

I think in such cases you need to improve SNR of the J-test measurement, usually things show up when you keep improving that part. But my take on this clock thing is that after changing a clock you must first measure it to check if it performs at least as well as before changing it. Because in many cases the difference is measurable. If it is not measurable and you still hear differences, that is all fine. But at least you know it didn't get objectively worse. I'd call it "sanity check". Because nice numbers on a datasheet/specs doesn't automatically mean that it will perform nice in the particular use case in real world! IOW, you cannot assume the change is unmeasurable.

 

For example using some fancy 10 MHz "atomic clock" as reference means that you need to have DPLL to generate the actual audio clock frequencies from that. And designing a good DPLL is not so easy, so it can easily go bad at that point. Another aspect is that you usually get best performance when you don't have clock dividers, but the clock is really running straight at the same frequency as the conversion section. But you probably cannot do that with any DAC chip, you need to go for a discrete DAC to do that. So these are some of the clocking related things, but unrelated to the clock module itself.

 

J-test24 is good for measuring clock performance, because you have a synchronous undithered test tone and rest of the spectrum is "black hole".

 

3 hours ago, barrows said:

I also am not trying to suggest that we cannot measure these things, just that we do not measure them all with the current set of measurements traditionally done-I think we need better measurements, I am really interested in what those might be.

 

Traditional set of measurements is not good for example for what I do with filters and modulators. I do also standard set, but in addition I do bunch of other measurements that are focused specifically on filter and modulator performance.

 

[Miska]

2 hours ago, Jking said:

The Mutec with Rubidium clock use to be popular on these pages.  

 

 

 

Now you'll only need to check how it performs in real world when used as clock for DAC instead of the built-in crystal...

 

Challenge is still the DPLL needed to generate useful clock from 10 MHz instead of internal crystal that is already running at needed frequency.

 

[Miska]

1 hour ago, barrows said:

Clearly you appear to believe that current measurements are enough, this is where we depart a bit.

 

Current measurements already show differences and you hear differences, you just need to find the correlation...

 

1 hour ago, barrows said:

So your take away is that humans, apparently, can hear things which most research says we cannot.  I certainly cannot hear a single tone at 20 kHz in room, but you are suggesting that perhaps I can hear something at that frequency and higher?

 

You cannot hear sine above 20 kHz, but it is not same as having over 20 kHz component as part of transient rise time. It is too simplistic to assume decomposed spectrum components alone would work the same way as the combination.

 

I personally have problem with metal dome tweeters that have resonance frequency somewhere between 22 and 28 kHz. Listening to such in long term is like listening to dentist ultrasonic teeth cleaning tool. If you know what I mean, that cleaning tool sounds just like telephone line modem handshake from the 80's-90's. Of course the tweeter is not as bad, but you get sort of headache kind of annoying feeling in back of your head. And it is much worse if the DAC has leaky filters. You often notice this at shows/fairs where you suddenly enter listening room and get a feeling of the sound in couple of first seconds before you even register what you are listening to.

 

1 hour ago, barrows said:

I do hear the difference if a swap to an XO with a bit better (6 dBc/Hz or more) low frequency phase noise performance

 

I don't have any doubts about that in general. My doubts are more like swapping clocks to commercial products with messy wires and stuff. That's why I'm after sanity check measurements just like for every other aspect too. Always measure first to sanity check that things didn't at least get worse...

 

Certainly those things matter (given good PCB layout design, materials and everything else), no doubt about that. My message is mostly that there is a lot to improve in terms of PCB layout design and other circuit design, just replacing a clock module is likely doing smaller difference than a change in PCB layout or PSU design. Or even capacitor selection. Unless you are starting with a really poor clock. Not to even forget how much design of the analog LPF section matters!

 

 

 - A person who has pile of Crystek CCHD-957's at 22.5792 & 24.576 MHz in stock and looks into 10 Hz phase noise figure...