In the thread (now closed) "The battle of the DACs, comparison of sound quality between some DACs" there was some discussion (or debating) about the impact of close-in phase noise on DAC output.
One way to measure the phase noise is to look at the noise skirts in FFT spectrum as JosephK did here: https://www.diyaudio.com/community/...jitter-crystal-oscillator.261651/post-6710439. This requires a very small bin size. I decided to make similar measurements with REW albeit the FFT size in REW is limited to 4M. But anybody with REW, DAC and ADC can repeat these measurements.
I used my own boards for the measurement:
First 11.025k sine of -1.5dBFS at 44k1. This uses 22.5792MHz ECS-2520MVLC clock having phase noise in datasheet as -89dBc/Hz @10hz (https://www.mouser.fi/datasheet/2/122/ECS_2520MVLC-1903379.pdf).
Note the horizontal scale (+/-10Hz). Not bad for a $1 clock.
Next up 12k sine of -1.5dBFS @48k. This uses a 24.576MHz NZ2520SDA clock with presumably lower close-in phase noise.
Clearly better but whether or not audibly better is another story.
Besides clock there are other factors that affect phase noise. One is Vref. In my AK4490 DAC Vref is fed by LT3042 regulator (separate L/R).
Here is the same measurement as above but with slightly altered Vref implementation.
Clearly much worse and a much bigger impact than clocks.
The only difference between graphs 2 and 3 was that Cset capacitor of LT3042 was changed from 22uF/15V tantalum polymer to 22uF/25V X5R. The result is understandable as Cset affects the 1/f noise of LT3042. Input and output capacitors of LT3042 are not critical for Vref phase noise.
One way to measure the phase noise is to look at the noise skirts in FFT spectrum as JosephK did here: https://www.diyaudio.com/community/...jitter-crystal-oscillator.261651/post-6710439. This requires a very small bin size. I decided to make similar measurements with REW albeit the FFT size in REW is limited to 4M. But anybody with REW, DAC and ADC can repeat these measurements.
I used my own boards for the measurement:
- AK4490 DAC
- ES9822PRO ADC
- USBI2S bridge
First 11.025k sine of -1.5dBFS at 44k1. This uses 22.5792MHz ECS-2520MVLC clock having phase noise in datasheet as -89dBc/Hz @10hz (https://www.mouser.fi/datasheet/2/122/ECS_2520MVLC-1903379.pdf).
Note the horizontal scale (+/-10Hz). Not bad for a $1 clock.
Next up 12k sine of -1.5dBFS @48k. This uses a 24.576MHz NZ2520SDA clock with presumably lower close-in phase noise.
Clearly better but whether or not audibly better is another story.
Besides clock there are other factors that affect phase noise. One is Vref. In my AK4490 DAC Vref is fed by LT3042 regulator (separate L/R).
Here is the same measurement as above but with slightly altered Vref implementation.
Clearly much worse and a much bigger impact than clocks.
The only difference between graphs 2 and 3 was that Cset capacitor of LT3042 was changed from 22uF/15V tantalum polymer to 22uF/25V X5R. The result is understandable as Cset affects the 1/f noise of LT3042. Input and output capacitors of LT3042 are not critical for Vref phase noise.
Phase noise of clocks probably should be compared at the same frequency. IIRC, phase noise measurements posted by @deleted_member showed that 24Mhz was visibly worse than 22Mhz.
Regarding NDK SDA series clock modules, I was informed by Jocko Homo that phase noise varies quite a bit from module to module. He told me that about 1 in 10 was exceptionally good, but it wasn't worth the labor to measure 10 in order to find one good one.
Regarding potential phase noise skirt audibility, when playing a recording of a 100-piece orchestra there are perhaps 1,000 or more frequencies present at once. The skirts of each frequency add up to raise the total noise floor in a way that is correlated with the audio signal. According to ESS, modulation of the noise floor is far more audible than if the noise floor is constant.
Would agree that Vref (AVCC) power quality is critical to perceived SQ. An experiment with an ADM7150 regulator procured from Twisted Pear and used to power AVCC on an Iancanada ES9038Q2M dac board showed that resistor loading of the regulator improved perceived SQ somewhat. A different regulator improved SQ further.
Regarding NDK SDA series clock modules, I was informed by Jocko Homo that phase noise varies quite a bit from module to module. He told me that about 1 in 10 was exceptionally good, but it wasn't worth the labor to measure 10 in order to find one good one.
Regarding potential phase noise skirt audibility, when playing a recording of a 100-piece orchestra there are perhaps 1,000 or more frequencies present at once. The skirts of each frequency add up to raise the total noise floor in a way that is correlated with the audio signal. According to ESS, modulation of the noise floor is far more audible than if the noise floor is constant.
Would agree that Vref (AVCC) power quality is critical to perceived SQ. An experiment with an ADM7150 regulator procured from Twisted Pear and used to power AVCC on an Iancanada ES9038Q2M dac board showed that resistor loading of the regulator improved perceived SQ somewhat. A different regulator improved SQ further.
Last edited:
Where has @deleted_member posted measurements of actual DACs? This was not about phase noise measurements of clocks.
If you feel that NDK SDA clock is not up to the task maybe you could show similar measurement from your own AK4499 DAC. Note also that the phase noise shown in my measurement is very similar to Joseph K's of dual AK4499 with TWTMC DRIXO 22,579MHz clock.
I have not made any claims regarding noise skirt audibility. These were just measurements to show the impact of phase noise to DACs.
To whatever extent dac skirts are a function of clock phase noise, skirts should be expected to correlate with measured phase noise, no?
Didn't say NDK SDA are not up to the task. Said you get the luck of the draw. Some are better than others, is all.
Understand you did not make claims as to audibility. Thought some readers might be interested to discuss that possibility.
Didn't say NDK SDA are not up to the task. Said you get the luck of the draw. Some are better than others, is all.
Understand you did not make claims as to audibility. Thought some readers might be interested to discuss that possibility.
This a guess or based on some analysis?
And yet again that ESS powerpoint gets rolled out. Is there no evidence anyone has tried to replicate this in any of the AES papers?
Each instrument has a set of harmonics to begin with. Then each instrument changing over time (they are non-steady state signals), which implies the existence of additional frequencies. 1,000 seems like it could be roughly in the ball park. Could be more.
Not that I am aware of. However the world has changed, government doesn't fund public research like it once did. Similar problem noted by @lrisbo who IIRC said to the effect that Purifi have done some casual collaboration with university researchers but they don't have the resources to produce publication quality human perceptual research. I find that quite easy to believe, personally.
DAC noise skirts are a function of clock phase noise but also other things such as Vref noise. IMO my measurements point this out. What these measurements also show is that ultra-low close-in phase noise clock is easily trumped by e.g. noise of Vref. A good example of this is Topping D90 where the Vref noise is very high due to poor design (mediocre Vref regulator).
It is quite odd that you readily accept phase noise measurements of clocks alone as some sort of proof but seem eager to dismiss phase noise measurements made of actual circuits using those clocks.
I didn't say that I dismiss anything 
Your measurement of a pathologically bad Vref supply shows it can be a problem too. No disagreement there. IMHO Vref quality is a critical part of good dac design. Clocks matter as well. What I have said is that if your dac is otherwise good enough then you will hear the effects of clock quality. A good Vref supply is part of being 'otherwise good enough.'
Your measurement of a pathologically bad Vref supply shows it can be a problem too. No disagreement there. IMHO Vref quality is a critical part of good dac design. Clocks matter as well. What I have said is that if your dac is otherwise good enough then you will hear the effects of clock quality. A good Vref supply is part of being 'otherwise good enough.'
Last edited:
Don't jump into conclusions. The Vref in my last measurement is not pathologically bad. The result is still miles ahead of e.g. Topping D90. Before making any quantitative (or qualitative) assessments try measuring your own AK4499 DAC and compare the results.
I'm more interested in how it sounds than how its FFT looks. The clocking of my custom AK4499 dac was improved until related SQ reached a point of diminishing returns. It took a lot of experimental work to get the clocking to that point. I listed some of the factors in the now closed thread you referred to. Before going to all that trouble various other people's clocking designs were tried. None of those was good enough except for Andrea's. His clocks may be better than really needed for my particular dac, but the important thing is they meet or exceed 'good enough.' Because the effects of clocks in that dac are audible it would seem to imply that its low noise discrete Vref regulator is reasonably good (although I think it could be further improved for reasons other than clock effect audibility).
Regarding Topping D90 sound, I am on record saying it has SQ problems, and that it would have been the worst sounding dac at the listening session.
Regarding Topping D90 sound, I am on record saying it has SQ problems, and that it would have been the worst sounding dac at the listening session.
Last edited:
I expect it sounds pathological compared to something better. Again, I am not that interested in how FFTs look. Their correlation with SQ is quite limited. Nobody can predict exactly how a dac will sound from such measurements. Two dacs can measure almost the same and still sound quite different. It means the measurements are of limited utility.
Is it? The X5R capacitor's capacitance will be somewhat reduced due to its non-linearity, but only slightly at 20 % of the rated voltage. Is it the microphony that makes one 22 uF capacitor so much worse than the other? If so, then it seems to contradict @gerhard's statements, he uses X5R capacitors in sensitive equipment that gets launched into space without running into microphony issues.
I was a bit surprised of this myself as just measuring the noise at LT3042 output did not reveal any difference.
And naturally the real-life progression was the opposite. First I had the X5R on Cset. Phase noise looked worse than in my ES9038Q2M dac. After some iterations I found that the cure is to use a tantalum polymer at Cset (which I already had on ES9038Q2M).
And naturally the real-life progression was the opposite. First I had the X5R on Cset. Phase noise looked worse than in my ES9038Q2M dac. After some iterations I found that the cure is to use a tantalum polymer at Cset (which I already had on ES9038Q2M).
Last edited:
Very interesting findings!
I had the same thoughts as MarcelvdG, including the possible microphony. Do you have any low frequency noise sources present, like fans etc.?
I am not an expert on space applications, but at least there shouldn't be much acoustic noise in that environment
But perhaps vibrations (if present) from internal stuff can couple to the capacitors?
For the tantalum I assume that the leakage current could be a concern, since it could lead to errors in the output voltage of the LT3042 due to the current source principle. For level precision this could be a concern. Have you encountered any problems in this regard?
Did you run the ES9038Q2M in the asynchronous or the synchronous mode when you made the noise skirt comparison with the AK4490?
I had the same thoughts as MarcelvdG, including the possible microphony. Do you have any low frequency noise sources present, like fans etc.?
I am not an expert on space applications, but at least there shouldn't be much acoustic noise in that environment
But perhaps vibrations (if present) from internal stuff can couple to the capacitors?For the tantalum I assume that the leakage current could be a concern, since it could lead to errors in the output voltage of the LT3042 due to the current source principle. For level precision this could be a concern. Have you encountered any problems in this regard?
Did you run the ES9038Q2M in the asynchronous or the synchronous mode when you made the noise skirt comparison with the AK4490?