Discussion split from here - https://www.diyaudio.com/community/threads/discrete-fpga-dac-project.407618/
Last edited:
Marcel,
IIUC, in theory for a shared clock the timing error cancels. However, each ADC/DAC device also has its own noise contribution. There may also be some clock skew arriving at the devices.
Moreover, modern ADCs and DACs have some processing time. Its not clear if the processing time cancels out. IOW, the dac outputs a sample at t=0, it passes through the output filter with a small delay, passes through a DUT, passes through an ADC input filter, then is oversampled with whatever clock jitter there is at that time. Its not clear if all that stuff cancels out exactly. To some extent some of the processing may attenuate some effects of jitter, but maybe less so for close in phase noise.
IIUC, in theory for a shared clock the timing error cancels. However, each ADC/DAC device also has its own noise contribution. There may also be some clock skew arriving at the devices.
Moreover, modern ADCs and DACs have some processing time. Its not clear if the processing time cancels out. IOW, the dac outputs a sample at t=0, it passes through the output filter with a small delay, passes through a DUT, passes through an ADC input filter, then is oversampled with whatever clock jitter there is at that time. Its not clear if all that stuff cancels out exactly. To some extent some of the processing may attenuate some effects of jitter, but maybe less so for close in phase noise.
Last edited:
There may be some cancellation but the above type of phase noise measurements are not meant to give absolute values. Rather they should be used for comparisons which makes the impact of cancellations more or less nonrelevant.
It is possible to process an asynchronously acquired test signal such that it doesn't have to be windowed. IIRC, Scott Wurcer once described doing it. It involves resampling the acquired sequence to make it synchronous with the acquired test waveform. Then truncation of the resampled sequence so as to prevent leakage. So the usual costs of resampling are there, but at least no windowing needed.
Last edited:
Close-in phase noise is nothing else than slow random clock frequency variations, and you try to make your measurement insensitive to small clock frequency inaccuracies.
I don't understand why you dislike Hann windows. OK, you need twice the number of samples for the same frequency resolution, but you don't have to worry anymore about asynchronous signals and clocks and about the fact that anything involving sigma-delta modulation is not exactly periodic.
Actually, its rather difficult to verify that claim by measurements for SOA clocks. We don't know the time-domain waveform of the clock phase deviation from ideal. It may be speeding up and slowing down at multiple frequencies at once and maybe not even for a whole cycle at once. Noise by its nature is often impulsive at some level, even if its averaged effects appear smoother.
So are you now excluding Vref noise from close-in phase noise?
I have not said that. With asynchronous clocks which clock is being measured? ADC's or DAC's? Another thing is that with asynchronous clocks and FFT windowing the differences are mostly hidden as not all spectral leakage is removed. With asynchronous clocks most DACs have similar noise skirts. Only with synchronous clocks the differences become apparent.
You can demodulate clock phase noise, that' true, but the demodulator will have its own noise. Last I heard years ago, the rack of equipment to do it with SOA clocks cost about $50k. Don't know if its still practical with today's best clocks.
However, for things like TIE analysis you need a clock better than the DUT clock.
Mostly what we know how to measure for SOA clocks is statistically averaged phase noise versus offset frequency. That's not the same as the time domain waveform of phase deviation. You can't get there from here, type of thing.
Also, isn't it true that what causes some phase noise are imperfections in crystals, surface contamination, their electrical contacts, etc. IOW, there are various physical causes? If there are enough we could point to the usual understanding of 1/f noise.
However, there is also the point Bart Kosko makes about bell curves in statistics: https://www.edge.org/response-detail/11715
Last edited:
BTW regarding terminology "phase noise" is a bit vague. IMO better terminology is phase modulated (PM) noise and amplitude modulated (AM) noise. The measurements of noise skirts in this thread contain both PM and AM modulated noise. I have not seen pure PM noise measurements of audio DACs.
Here is a good paper on phase noise basics and measurements:
https://www.ab4oj.com/test/docs/20180720_KEE7_PhaseNoise.pdf
Here is a good paper on phase noise basics and measurements:
https://www.ab4oj.com/test/docs/20180720_KEE7_PhaseNoise.pdf
By trying different signal frequencies, for example. Or maybe by either synchronizing the clocks or using the same voltage reference for the ADC and DAC.
I have on occasion tried different frequencies but phase noise remains the same (with or without synchronous clocks). My assumption is that AM noise dominates. Having same Vref on both ADC and DAC is not feasible in practice.
Regarding what's noise and what is signal, in an FFT a spectral line is typically considered a signal. Anything spread out over a few or several bins looks more like noise. However, what looks like noise in an FFT is not necessarily noise in the time domain. It could be perfectly deterministic signal in the time domain, such as a frequency sweep. Its just not PSS. Any non-PSS signal tends to look like noise on an FFT.
Yes, if long averages are used. Aperiodic signals can be seen in snapshots or even with averages if peaks are calculated.
It's not surprising, not to me anyway, that amplitude noise tends to dominate over phase noise. On the well-tempered master clock thread, I once calculated the sideband noise due to the phase noise of a reasonable crystal oscillator (not one of Andrea's super crystal oscillators) and compared that to sideband noise measurements of @Joseph K The phase noise was way below the total, because in his case, amplitude noise due to the voltage reference was also very dominant.
However, for soundstage imaging its not clear that the ear necessarily mistakes amplitude errors for timing errors, and or vice versa. IOW, just because Vref noise and phase noise are closely correlated with an audio signal in a phase-free FFT, doesn't mean they sound the same in the time domain.