abzug,
can't wait to see the results of your project.. sounds interesting. I have considered for some time to buy an evaluation board from anagram but what I need is just the scrambler dsp and maybe the q5 upsampler in order to be able to roll the dac chips I have.
What materials do you recommend for studying in regard of paralleling delta-sigma DACs?
can't wait to see the results of your project.. sounds interesting. I have considered for some time to buy an evaluation board from anagram but what I need is just the scrambler dsp and maybe the q5 upsampler in order to be able to roll the dac chips I have.
What materials do you recommend for studying in regard of paralleling delta-sigma DACs?
Although the anagram looks to be a better implementation, I have to say that at this moment I don't have the knowledge to code a dsp nor the resources, especially as I am working on a PC audio interface.
However I would like to try a parallel configuration for the ak4396 and ad1955 dac's I have.
How should I proceed with the input? simply feeding all the dac chips with the SDATA, LRCK, SCLK and MCLK would suffice?
Also, although I understand the the way the anagram version works, I don't really understand how in the case of simply paralleling the dac chips, (without the "scrambling" technology) the linearity increases. Some insight on this could help me a lot.
By the way, what do you think about this: 10 fully differential 32-bits DAC ?
Thank you!
However I would like to try a parallel configuration for the ak4396 and ad1955 dac's I have.
How should I proceed with the input? simply feeding all the dac chips with the SDATA, LRCK, SCLK and MCLK would suffice?
Also, although I understand the the way the anagram version works, I don't really understand how in the case of simply paralleling the dac chips, (without the "scrambling" technology) the linearity increases. Some insight on this could help me a lot.
By the way, what do you think about this: 10 fully differential 32-bits DAC ?
Thank you!
Hello,
The image bellow shows in which way the anagram solution should outperform the conventional paralleling. It seems that even with 16 devices in parallel, the distorsion performance is not as good above 4K.
The out of band performance of the anagram solution could maybe affect the 3-20K region as well?
And what do you think about the output stage designed for the AD1955, bellow?
The image bellow shows in which way the anagram solution should outperform the conventional paralleling. It seems that even with 16 devices in parallel, the distorsion performance is not as good above 4K.
The out of band performance of the anagram solution could maybe affect the 3-20K region as well?
An externally hosted image should be here but it was not working when we last tested it.
And what do you think about the output stage designed for the AD1955, bellow?
An externally hosted image should be here but it was not working when we last tested it.
Simple paralleling of DACs shouldn't do much for their linearity, I don't think. The big improvement to be had is a 3dB reduction in output noise per doubling of devices (when the noise between devices is uncorrelated).
When part of the distortion of a given device is due to process imperfections (like component mismatch or whatever), and is therefore different in each one, then these effects will sort've average out when lots of them are paralleled. However, the parts of the distortion that are inherent to the design of the DAC itself, and are therefore the same in each one, won't be affected by paralleling. That is, I think, what is visible in that distortion graph you posted - the LF THD+N is dominated by noise, and the HF THD+N is dominated by distortion. Paralleling only helps the former.
What anagram seem to be trying to do is to dither out the DAC's nonlinearities. This might help to some extent, although I think rigorous decorrelation of errors is... ambitious. The measurements that have been posted so far don't really say much of anything, and I haven't been able to find any more. I doubt this process magically makes any old DAC distortionless, though.
By the way, what is it about upsampling that Anagram do better than anyone else? Upsampling well isn't hard. Their website seems to be down...
When part of the distortion of a given device is due to process imperfections (like component mismatch or whatever), and is therefore different in each one, then these effects will sort've average out when lots of them are paralleled. However, the parts of the distortion that are inherent to the design of the DAC itself, and are therefore the same in each one, won't be affected by paralleling. That is, I think, what is visible in that distortion graph you posted - the LF THD+N is dominated by noise, and the HF THD+N is dominated by distortion. Paralleling only helps the former.
What anagram seem to be trying to do is to dither out the DAC's nonlinearities. This might help to some extent, although I think rigorous decorrelation of errors is... ambitious. The measurements that have been posted so far don't really say much of anything, and I haven't been able to find any more. I doubt this process magically makes any old DAC distortionless, though.
By the way, what is it about upsampling that Anagram do better than anyone else? Upsampling well isn't hard. Their website seems to be down...
I want to find out when I implement it, but the problem is I don't have access to any good measuring equipment, and I can't exactly drop $7K to $20K on a dual domain Audio Precision. Of course, if I can hear an improvement in blind testing that would be reason enough to use it.SunRa said:
What they're doing seems to be similar to what the AD1955 already does internally (by randomizing the mapping of the thermometer code to dynamic elements) across multiple DACs. That may be also why there seems to be no patent on it.
Forget the site; look at their patent. Go to Google Patents and type:Wingfeather said:
inassignee:"anagram technologies"
To an extent it has to do with efficiency of the implementation, and they do claim some filtering improvement. It's a two pass filter with time domain spline interpolation in the second pass.
I can acces it... maybe it was temporarily.
This technology maybe: adaptive time filtering
What do you think about this "second generation" of paralleled devices
from Accuphase? It's not just the usual way of paralleling devices.
"After volume processing, the timing of the 1-bit signal is shifted progressively in increments of 177 nanoseconds (about 5,600,000th of a second), using high-performance circuitry. The resulting delayed signals are sent to multiple D/A converters (eight per channel in the DC-801) for conversion into analog form. The converter outputs are then summed. Signal components with precipitous changes, i.e. high frequencies are sequentially averaged, so that the entire circuit acts as a high-cut filter that removes unwanted noise components in the high frequency range."
The link is
here
And a picture:
An externally hosted image should be here but it was not working when we last tested it.
Thanks for that.
I'm surprised though that they are able to patent what essentially seems to be spline interpolation. I guess it's the specifics that they are trying to protect. If you're into ASRCs then Q5 seems like a good system - although, anyone can come along and do quadratic B-spline interpolation (Newton polynomial interpolation of varying orders also works well, though I don't know about the relative computational effort) and get a similarly good result.
I think it should be remembered that for integer upsampling ratios, a garden-variety FIR interpolator will still be better.
Which bit is bad?
Wikipedia has good articles on interpolation methods. They involve taking a selection of audio samples (say, four adjacent samples), and reducing them into a (third-order, in this example) polynomial equation in time that describes them. Once you have this, you can compute the value of the equation at any arbitrary point in time to work out the value of the signal at that point. It's quite useful, but the problem is that these four samples don't have all the information that was in the original signal and so the polynomial will only be an approximation to it. Using a bigger selection of samples will get you a higher-order polynomial which will be closer to the input signal. But you can never get there exactly.
Bezier-splines and Newton Forward Difference are just two of the ways of converting the audio samples into a mathematical function.
If one wants to go this route, a better way is to take advantage of the mechanics of digital audio and use a sinc function centred on each sampling instant and weighted to the sample's value. If you then sum the contributions of every sample's sinc function at a specific point in time then you get the value of the signal at that time. The problem is that unless you do it for *all* the samples then it's again only another approximation. Also, calculating the value of sinc is a lot of effort. So really, this method is only good for offline resampling. However, if you do do it for all samples then the result is perfect.
For integer upsampling ratios (like 44.1kHz->88.2kHz, etc), none of this nonsense is required. A simple FIR oversampler doesn't approximate the signal at all, it maintains it perfectly. What it does do is add spectral repeats of it at higher frequencies, but these can be filtered out with the FIR filter. If the filter has a stopband attenuation greater than the dynamic range of the word-length being used for the signal, then the images are essentially removed, and the interpolation is essentially perfect.
Wikipedia has good articles on interpolation methods. They involve taking a selection of audio samples (say, four adjacent samples), and reducing them into a (third-order, in this example) polynomial equation in time that describes them. Once you have this, you can compute the value of the equation at any arbitrary point in time to work out the value of the signal at that point. It's quite useful, but the problem is that these four samples don't have all the information that was in the original signal and so the polynomial will only be an approximation to it. Using a bigger selection of samples will get you a higher-order polynomial which will be closer to the input signal. But you can never get there exactly.
Bezier-splines and Newton Forward Difference are just two of the ways of converting the audio samples into a mathematical function.
If one wants to go this route, a better way is to take advantage of the mechanics of digital audio and use a sinc function centred on each sampling instant and weighted to the sample's value. If you then sum the contributions of every sample's sinc function at a specific point in time then you get the value of the signal at that time. The problem is that unless you do it for *all* the samples then it's again only another approximation. Also, calculating the value of sinc is a lot of effort. So really, this method is only good for offline resampling. However, if you do do it for all samples then the result is perfect.
For integer upsampling ratios (like 44.1kHz->88.2kHz, etc), none of this nonsense is required. A simple FIR oversampler doesn't approximate the signal at all, it maintains it perfectly. What it does do is add spectral repeats of it at higher frequencies, but these can be filtered out with the FIR filter. If the filter has a stopband attenuation greater than the dynamic range of the word-length being used for the signal, then the images are essentially removed, and the interpolation is essentially perfect.
Sinc is infinite in extent. This sort of filter is invalid unless its length is as long as the piece of music you're playing, or at least long enough that the sinc value at your cutoff points is low enough that it cannot affect the least significant bit before dithering--and around 40 bit is considered good in floating point (Anagram uses a Blackfin which is fixed point and has an 80-bit accumulator). That's a very long, extremely slow filter, that can only be done offline. And do you actually want a filter that's a brickwall in the frequency domain?Wingfeather said:
As for FIRs, they are not optimal either.
You are failing to account for the fact that the human ear is a dual-domain instrument, that analyzes sound in both frequency and time domain. We can hear individual clicks spaced at much less than 1/(20 kHz). It is incorrect to assume that linear phase filters like symmetric FIRs with their preringing are optimal. It's no coincidence that Wolfson Micro is researching filters which are between minimum and linear phase:
"An Ultra High Performance DAC with Controlled Time-Domain Response" by Lesso, Paul; Magrath, Anthony (119th AES Convention).
It's not a filter. It's representing the signal as a series of weighted sinc functions instead of a series of weighted rectangular pulses. It is long, and it is slow, and that's why I already said it's only suitable for offline processing. Going to 40-bit precision seems excessive, but even stopping at 24-bit is definitely going to have quite a latency. I suppose it did sound like I was saying sinc-interpolation was suitable for high-quality realtime operation, which I really didn't mean. Just, AFAIK it's the best that can be done for ASRC operations in theoretical terms.
The Blackfin is a 16-bit DSP with 40-bit accumulators.
That paper seems to be saying that it might be prudent to consider time-domain effects too, but that there's little to no evidence that it's a requirement. I think this:
is an appropriate quote. That said, evidence in either direction on this sort of thing is very hard to come by and I do personally agree that it's worth minimising things like pre-ringing where possible. FIR interpolators can be broken down into multiple 2x stages which eases requirements on filter length. After the first 2x stage, transition bands can also get larger which allow less pre-ringing for a given stopband attenuation. Given appropriate care, total system pre-ringing can be made very low.
You're right; I was thinking of the SHARC which, when used in fixed point, has an 80 bit accumulator.Wingfeather said:
http://www.stereophile.com/reference/106ringing/index.html
I still don't understand the insistence on FIR filters. Even if you want a linear phase filter, why not use say a very high order linear phase Bessel IIR?
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.