AD1865 vs AD1955

Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.
Different animals. First is a multibit R-2R type with THD+N of -90dB, the second is "multibit sigma-delta" with THD+N of -110dB. Usually that means there are a few multibits (usually 4-10, they don't say how many here), each modulated by a sigma-delta signal.

On paper the second is definitely better and as proof... it is used in a lot of high-end devices.
Some people say that the pure multibit (like your first one) sounds "better" than a noise-shaped delta-sigma.
To me... it's a close call. I have both types and I listen mainly to a 18 bit multibit too because I think it sounds slightly better while playing CD signals.
But I didn't conduct extensive tests, when I have a hi-res format or a SACD, I am using the sigma-delta.
 
Last edited:
Ok here are the two plots I posted there - I realize now you have to sign up to see them.

What's hidden (because these are FFTs) is any short-term changes in the noise floor. FFT is just showing us the average noise over the acquisition window, but we know that the noise out of an S-D loop is instantaneous level dependent. ESS's Sabre datasheet shows that - the noise varies with DC level.

The noise spectrum is also dependent on the signal level - notice how it slopes upward with zero signal, but ostensibly remains flat with a full-scale sinewave present. Thus we can expect the timbre of the noise to be a dynamic thing.

If you think multibit has the same issue, then post up the plots to show it.
 

Attachments

  • ad1955-0dB.png
    ad1955-0dB.png
    35 KB · Views: 780
  • ad1955-low.png
    ad1955-low.png
    34.2 KB · Views: 758
At -140..-130dB it is unlikelly to be heard.

You're making the common misinterpretation of an FFT. The data points represent noise voltage into a very narrow bandwidth - much narrower than the critical bands in the ear.

THD are at -112..115dB and they will cover any noise (the result THD+N is at -110dB).

How do you know?

Let's not hijack this topic, maybe we can make another one for this.

This is germane - the OP is asking for differences in sound between two ADI parts. I hear that the AD1955 is less dynamic and in the absence of any other reasonable hypothesis, I put it down to noise modulation. Those plots are the smoking gun. What more do you want? :D
 
If I understand what you're trying to say abraxalito, you're basically pointing out that FFT analyses are done over an average rather then being instantaneous right? And because of this, any transient increases in noise, due to the D-S modulation, will go unnoticed.

Surely though, if one were to measure the THD @ say 20kHz, any large amount of instantaneous noise would show up as a degradation in linearity as increase in THD? The noise would effectively modulate the sine wave in a non linear fashion and show up in the final result.

A sine wave might be a boring old wave-form, but it is one that is predictable and well known. If any part of the output sine wave deviated, due to some spurious noise, from what was expected then surely something should show up.

If this change in noise level, due to the signal level changing, is that quick though, surely it will be filtered out by any low-pass following the DAC? Isn't this the point of using noise shaping in the D-S DAC anyway or have I missed the point?
 
If I understand what you're trying to say abraxalito, you're basically pointing out that FFT analyses are done over an average rather then being instantaneous right? And because of this, any transient increases in noise, due to the D-S modulation, will go unnoticed.

Pretty much yep. FFTs are typically done with 1k points or more often more. 1k points at 44k1 means about 23mS of signal. As a result, the noise looks a lot lower on the plot (-140dB) because its averaged over that time. Alternatively another way to think of it is an FFT is a bank of matched filters, each with a narrow bandwidth. Each of those filters is giving its result (-140dB) but to get the noise in the total bandwidth those contributions must be summed (power-wise if uncorrelated) together.

Surely though, if one were to measure the THD @ say 20kHz, any large amount of instantaneous noise would show up as a degradation in linearity as increase in THD? The noise would effectively modulate the sine wave in a non linear fashion and show up in the final result.

Why would this be likely to show on 20kHz THD and not at 1kHz? Noise is not harmonic distortion - its not generally correlated with the signal.

A sine wave might be a boring old wave-form, but it is one that is predictable and well known. If any part of the output sine wave deviated, due to some spurious noise, from what was expected then surely something should show up.

What we have is evidence that the noise floor is changing - between a full-scale signal and a very low level (or absent) signal. For 67% of the time the sinewave's (assuming its 0dBfs) instantaneous amplitude is at or above -6dB. Its spending almost no time at all down at the -100dB level. But music is quite a different beast - with much higher crest factor than a sine. It spends more time down at lower instantanous levels.

If this change in noise level, due to the signal level changing, is that quick though, surely it will be filtered out by any low-pass following the DAC? Isn't this the point of using noise shaping in the D-S DAC anyway or have I missed the point?

If the noise comes and goes a few times within a 23mS window, why would that be filtered out? Its not a high bandwidth modulation of the noise we're talking about.
 
If the noise comes and goes a few times within a 23mS window, why would that be filtered out? Its not a high bandwidth modulation of the noise we're talking about.

Because it is shaped noise with dither? Noise "heavy" in high-frequency, out-of-band components that get filtered. It is not thermal white noise, it's pre-calculated.
 
Last edited:
Why would this be likely to show on 20kHz THD and not at 1kHz? Noise is not harmonic distortion - its not generally correlated with the signal.

I picked 20kHz due to your previous comment about D-S DACs seeming to dynamically compress everything. As 20kHz has close to the fastest rise time that we'd expect a normal 44.1k system to be able to reproduce then if a D-S were really screwing with the dynamics, surely it would show up better, if at all, at 20k then 1k.



What we have is evidence that the noise floor is changing - between a full-scale signal and a very low level (or absent) signal. For 67% of the time the sinewave's (assuming its 0dBfs) instantaneous amplitude is at or above -6dB. Its spending almost no time at all down at the -100dB level. But music is quite a different beast - with much higher crest factor than a sine. It spends more time down at lower instantanous levels.

Surely this is good for the music then as it implies less noise will be produced when operating with music, rather then with the sine. I mean I've noticed a few times when measuring DACs that the noise increases when the signal level goes up. As far as I remember I measured the same for TI's digital input class D amps that also use noise shaping.

I mean I see this as being rather simple, the system noise goes up as it is asked to work closer to it's upper limits. I don't actually see any problem with this providing the noise isn't doing what you're saying it's doing.



If the noise comes and goes a few times within a 23mS window, why would that be filtered out? Its not a high bandwidth modulation of the noise we're talking about.

Right.

So if some transient comes along and causes a burst of noise to go along with it (and this is what I am taking to be the cause of why you think D-S dulls dynamics), then the noise will be completely random and be within the audible band. Surely this should be measurable as an instantaneous change in the noise floor, or better yet, why not ask the DAC to reproduce 1Hz and actually watch as the noise floor changes with the signal level. Or, are you saying that this only really occurs if the DAC is asked to reproduce a wave with a fast rise time?
 
I picked 20kHz due to your previous comment about D-S DACs seeming to dynamically compress everything. As 20kHz has close to the fastest rise time that we'd expect a normal 44.1k system to be able to reproduce then if a D-S were really screwing with the dynamics, surely it would show up better, if at all, at 20k then 1k.

When I talk about lack of dynamics, I'm not saying the transients are absent, rather I'm saying they lose subjective impact. One way to achieve this is to add correlated noise to them. This way the perceived dynamic range is reduced. I can't hear anything much near 20kHz myself and music with a lot of HF isn't required to hear the loss of dynamics.

Surely this is good for the music then as it implies less noise will be produced when operating with music, rather then with the sine.

I think that might be an assumption too far. We don't know just from those FFT plots at which instantaneous level the noise floor changes. Even assuming that it does have such a level rather than a continuous change.

Just thinking out loud here - suppose that the extra noise is only added when the sine (its instantaneous level) gets above -3dBfs. The high level FFT (the left plot) then represents the average noise floor but there could be peaks of 6dB above that average happening 50% of the time.

I mean I see this as being rather simple, the system noise goes up as it is asked to work closer to it's upper limits. I don't actually see any problem with this providing the noise isn't doing what you're saying it's doing.

The engineers at Dolby are probably the people who have the most experience with systems that exhibit noise floor modulation. A noise reduction processor is one such system. Those guys say noise floor modulation is audible, even in the presence of loud signals.

So if some transient comes along and causes a burst of noise to go along with it (and this is what I am taking to be the cause of why you think D-S dulls dynamics), then the noise will be completely random and be within the audible band.

Where are the transients in a sinewave test tone? By which I don't mean things with fast rise time but where the signal is spending a lot of time at low level and relatively short periods of time at high level. A sine is the opposite - most of its time at high levels and zipping through the lower levels very rapidly.

Surely this should be measurable as an instantaneous change in the noise floor

There's a big problem right there - we don't currently have widely accepted measurements for instantaneous noise. Noise is normally measured on average. We need a measurement which is simultaneously wide dynamic range and fairly wide bandwidth - wavelet probably will do it. But who's using wavelets in audio?

or better yet, why not ask the DAC to reproduce 1Hz and actually watch as the noise floor changes with the signal level.

Yes, this I believe has been done. I know I've tried it in the past, but without much success. The idea would be to try to correlate the noise with the instantaneous signal level.

Or, are you saying that this only really occurs if the DAC is asked to reproduce a wave with a fast rise time?

No, I don't think rise time has much (if anything) to do with it. But then again, I could be wrong. For sure noise modulation occurs when an S-D is asked to reproduce DC levels - ESS shows this in their datasheet.

<edit> I'll just mention that the effect of 'slowing' the dynamics might just as much be an issue for the I/V and filtering stage as it is for the DAC's S-D loop. When a high level signal comes along suddenly that's definitely going to be accompanied by a burst of higher frequency (out of band) noise - at much higher frequencies than for an I/V stage in a multibit DAC. It needs excellent linearity circuitry not to create in-band IM products as a result of the sudden increase in RF noise. Conversely at the lower signal levels (<30dBfs) then the OOB noise increases as the audio-band signal decreases because the overall output energy is roughly constant.
 
Last edited:
Just thinking out loud here - suppose that the extra noise is only added when the sine (its instantaneous level) gets above -3dBfs. The high level FFT (the left plot) then represents the average noise floor but there could be peaks of 6dB above that average happening 50% of the time.

If you're saying that these peaks in the noise floor are only a few dB above the average level then why do you think it is a a problem? Surely the noise floor in the acoustic will dominate.


Where are the transients in a sinewave test tone? By which I don't mean things with fast rise time but where the signal is spending a lot of time at low level and relatively short periods of time at high level. A sine is the opposite - most of its time at high levels and zipping through the lower levels very rapidly.

Why is this important? Sure the peak to average ratio might be low, but with a high frequency sine wave you're going from the largest minimum value to the largest positive value very quickly. I suppose nyquist theory, representing the highest frequency sine wave that the system can produce, might not be entirely accurate on the fastest signal the DAC might be asked to reproduce. Surely it can be asked to go from complete silence to full positive output within just one sample?

Yes, this I believe has been done. I know I've tried it in the past, but without much success. The idea would be to try to correlate the noise with the instantaneous signal level.

Wouldn't this indicate that it isn't really a problem? If the noise contribution was significant then maybe you'd have had more success?



No, I don't think rise time has much (if anything) to do with it. But then again, I could be wrong. For sure noise modulation occurs when an S-D is asked to reproduce DC levels - ESS shows this in their datasheet.

Does the duration that the DACs output is held to one specific value matter to the DAC? Surely an AC signal, from a DACs point of view, is just switching very quickly between discrete DC levels.
 
If you're saying that these peaks in the noise floor are only a few dB above the average level then why do you think it is a a problem? Surely the noise floor in the acoustic will dominate.

That's the wrong way around. I already hear there's a problem with the AD1955 - I'm following up leads to give possible explanations for what's causing it. I don't know what 'noise floor in the acoustic' means - there's a noise floor in the 16bit process (say around -93dB) and I think the problems of the AD1955 are probably going to be at least some of the time above this level. The dynamically changing noise floor is the problem, not the absolute level - a static noise floor at say -90dB I don't think I'd hear as a loss of dynamics.

Why is this important? Sure the peak to average ratio might be low, but with a high frequency sine wave you're going from the largest minimum value to the largest positive value very quickly. I suppose nyquist theory, representing the highest frequency sine wave that the system can produce, might not be entirely accurate on the fastest signal the DAC might be asked to reproduce. Surely it can be asked to go from complete silence to full positive output within just one sample?

I can't get your point from this - I suspect we're at crossed purposes :)

Wouldn't this indicate that it isn't really a problem? If the noise contribution was significant then maybe you'd have had more success?

I'm concluding from this that you're definitely an objectivist. I'm a subjectivist - I hear things and seek to understand why I hear them. So I trust what I hear more than a particular choice of measurement - if a measurement doesn't show the effect I'm hearing, try another one.

Does the duration that the DACs output is held to one specific value matter to the DAC? Surely an AC signal, from a DACs point of view, is just switching very quickly between discrete DC levels.

Once again, you've lost me :)
 
That's the wrong way around. I already hear there's a problem with the AD1955 - I'm following up leads to give possible explanations for what's causing it. I don't know what 'noise floor in the acoustic' means - there's a noise floor in the 16bit process (say around -93dB) and I think the problems of the AD1955 are probably going to be at least some of the time above this level. The dynamically changing noise floor is the problem, not the absolute level - a static noise floor at say -90dB I don't think I'd hear as a loss of dynamics.

The noise floor in the acoustic = the noise floor of your listening room. It will dominate, or should dominate significantly over any noise that the system produces. Any fluctuations of the noise floor in the digital to analogue process should be completely masked by your room.


Why is this important? Sure the peak to average ratio might be low, but with a high frequency sine wave you're going from the largest minimum value to the largest positive value very quickly. I suppose nyquist theory, representing the highest frequency sine wave that the system can produce, might not be entirely accurate on the fastest signal the DAC might be asked to reproduce. Surely it can be asked to go from complete silence to full positive output within just one sample?

I can't get your point from this - I suspect we're at crossed purposes :)

You were complaining that a sine wave isn't representative of real music. Music normally spending more of its time at a lower average level then a sine wave, then with occasional peaks up to 0dBfs lets say.

If we look at the maximum frequency sine wave that a DAC can reproduce then that should also show us the fastest rise time that any music signal could contain. Going directly from the trough to the peak of a full amplitude sine wave at this frequency should represent the worst case scenario for any transient, what came before this transient shouldn't be of any importance to the DAC.

The last sentence of what I said was a bit of a brain fart when I think about it, so please ignore it.



I'm concluding from this that you're definitely an objectivist. I'm a subjectivist - I hear things and seek to understand why I hear them. So I trust what I hear more than a particular choice of measurement - if a measurement doesn't show the effect I'm hearing, try another one.

We look at things from different ends then, but at least you want to actually be able to back up what you're hearing with a measurement.

Does the duration that the DACs output is held to one specific value matter to the DAC? Surely an AC signal, from a DACs point of view, is just switching very quickly between discrete DC levels.

Once again, you've lost me :)

This was in relation to what you would perhaps consider a transient effect. If you are trying to measure some sort of noise floor modulation then you should be able to do it with the DACs output held at a constant DC level. I was simply looking at it from the perspective of the DAC, it gives out a train of stepped voltages that could be looked at as a series of very quick changes in the DC level. From the point of view of the DACs internal operation and hence its noise output, does it care that the DC level has existed for only one sample or if if has existed for one hundred? If the DAC is going to do the same thing per sample, regardless of the previous sample, then you should easily be able to measure noise floor changes.

Having said this though, couldn't you simply use an analogue stage, after the DAC, that has a level of noise that vastly swamps the DACs own. Like this any noise floor modulation produced by the DAC should be completely dominated by the analogue stage. If it's modulation of the noise floor that is bothering you then surely that'd go a long way to removing it?
 
Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.