Yes that's correct. I like a more revealing DAC - one that doesn't put a veil between me and those (same) digits.
DCS Vivaldi
- Thread starter MylesBAstor
- Start date
In S-D DACs its noise modulation. In many R2R DACs, its glitching. That leaves a minority of multibit DACs to build revealing DACs with.
S-D? R2R? I googled SD DAC. God San Diego Animal Control. And I can't imagine why I'd need a DAC in my reel to reel.
Tim
Tim
Ah sorry I forgot to go slowly. OK, rewind, quarter speed ahead....
S-D means 'sigma delta' style DAC which relies on large oversampling ratios (64X and beyond) and low bit (say 6bits or fewer) DAC. It uses noise shaping to get decent audio band noise levels, but only the average noise turns out to be low - some of the noise is correlated with the signal. Example - PCM1792.
R2R refers to a type of DAC known as a ladder DAC where the individual bit weights (6dB steps) are made up of a series two-resistor networks - one with value R, the other with 2R. Example of this kind of DAC - PCM1704 (though its modified R2R).
S-D means 'sigma delta' style DAC which relies on large oversampling ratios (64X and beyond) and low bit (say 6bits or fewer) DAC. It uses noise shaping to get decent audio band noise levels, but only the average noise turns out to be low - some of the noise is correlated with the signal. Example - PCM1792.
R2R refers to a type of DAC known as a ladder DAC where the individual bit weights (6dB steps) are made up of a series two-resistor networks - one with value R, the other with 2R. Example of this kind of DAC - PCM1704 (though its modified R2R).
Very, very, VERY cool! Thanks Opus! Appreciate your taking the time to share your experience. What multibit DACs do you like? And, as you would expect....what digital equipment currently uses them?
Sorry, I should shut up. Opus is doing good, and the engineering jargon I spouted is explained more fully in the threads I linked in the tech forum.
Simplified (and thus less precise):
1. Random jitter from the clock source is unlikely to matter, it just adds low-level noise.
2. Some jitter is related to the signal or the clock, and that kind can add much higher distortion, though again rarely enough to matter in the real world.
3. Asynchronous DACs attempt to isolate the incoming clock and data from the signals used by the DAC component itself, the chip (or whatever) inside the DAC box. That allows designers to use a much cleaner clock source and provide data that is well time-aligned to the clock and avoid some types of jitter.
4. Long-term clock stability does not matter for listeners; it is important in a studio where multiple converters and other components must all work together and all at the same time to avoid jitter and other issues.
HTH - Don
p.s. I think I introduced various DAC architectures in my DAC threads in the tech forum, including R2R, unary, segmented, and delta-sigma. There are curves showing how DS DACs do noise shaping.
p.p.s. I say "delta-sigma" rather than "sigma-delta" because the modulator loop does differencing (delta) first then integration (sigma), and because a famous DS guy (Gabor Temes) told me in class decades ago it should be that way; the original paper was mis-translated SD instead of DS.
Can one 'sort of' generalize the sound of D S vs. R2R DAC's and if so what is generally the difference?
Hand-waving again, hopefully not too techie but no promises...
DS DACs perform noise shaping and use high-order modulators, high clock rates, and lower-resolution "DACs" in a feedback loop to provide very low in-band noise. Noise shaping can provide very high in-band SNR at the cost of very high out-of-band noise. The noise rises with frequency, unlike conventional circuits that generally exhibit flat or falling noise floors with frequency (to a point well beyond the scope of this topic). Other issues DS DACs may have are that the modulators may be subject to strange spurious artifacts (noise/distortion tones) caused by the digital filters when certain combinations of signal and clock frequencies are hit. Perhaps the biggest advantage of DS DACs from an audiophile POV is that their very high sampling rate means the output filter all DACs require to reject out-of-band images generated by the sampling process can be placed at a very high frequency. That means less roll-off, phase shift, ringing and such is added to the audio signal by the filter.
Conventional DACs like R2R designs generate a glitch (like a noise spike) every time the data changes, and that glitch can be large (for an R2R, half full-scale). Worst-case, think of the MSB toggling: when the MSB turns on at mid-scale, all the other bits turn off, and a large glitch is generated. Since this is right at mid-scale, the glitch impacts low-level signals the most because mid-scale of the DAC is 0 V (the center) of a +/- output signal. This glitch generates signal-dependent distortion.
A conventional DAC that is not oversampled requires a steep filter at half the sampling rate to reject images that appear starting at and beyond half the sampling frequency. A high-order filter is needed, and that means (usually) a lot of phase shift, ringing, and roll-off extending well into the audio band. Conventional DACs can be oversampled, of course, to mitigate the filter effects. For that matter, you can oversample and filter to improve the SNR of a conventional DAC, but without the noise-shaping modulator the benefit is fairly small.
Sonically, noise shaping (DS) converters have a noise floor that rises with frequency and thus may be more audible/annoying in the highs, and added non-harmonic tones may corrupt the sound. The glitches and filters of conventional DACs can put an edge on the highs and smear transients, and their noise floor tends to have a "raspier" or harsher sound (though added noise to dither the lsb mitigates that in both architectures). The audibility of any of these effects is debatable, natch.
HTH - Don
DS DACs perform noise shaping and use high-order modulators, high clock rates, and lower-resolution "DACs" in a feedback loop to provide very low in-band noise. Noise shaping can provide very high in-band SNR at the cost of very high out-of-band noise. The noise rises with frequency, unlike conventional circuits that generally exhibit flat or falling noise floors with frequency (to a point well beyond the scope of this topic). Other issues DS DACs may have are that the modulators may be subject to strange spurious artifacts (noise/distortion tones) caused by the digital filters when certain combinations of signal and clock frequencies are hit. Perhaps the biggest advantage of DS DACs from an audiophile POV is that their very high sampling rate means the output filter all DACs require to reject out-of-band images generated by the sampling process can be placed at a very high frequency. That means less roll-off, phase shift, ringing and such is added to the audio signal by the filter.
Conventional DACs like R2R designs generate a glitch (like a noise spike) every time the data changes, and that glitch can be large (for an R2R, half full-scale). Worst-case, think of the MSB toggling: when the MSB turns on at mid-scale, all the other bits turn off, and a large glitch is generated. Since this is right at mid-scale, the glitch impacts low-level signals the most because mid-scale of the DAC is 0 V (the center) of a +/- output signal. This glitch generates signal-dependent distortion.
A conventional DAC that is not oversampled requires a steep filter at half the sampling rate to reject images that appear starting at and beyond half the sampling frequency. A high-order filter is needed, and that means (usually) a lot of phase shift, ringing, and roll-off extending well into the audio band. Conventional DACs can be oversampled, of course, to mitigate the filter effects. For that matter, you can oversample and filter to improve the SNR of a conventional DAC, but without the noise-shaping modulator the benefit is fairly small.
Sonically, noise shaping (DS) converters have a noise floor that rises with frequency and thus may be more audible/annoying in the highs, and added non-harmonic tones may corrupt the sound. The glitches and filters of conventional DACs can put an edge on the highs and smear transients, and their noise floor tends to have a "raspier" or harsher sound (though added noise to dither the lsb mitigates that in both architectures). The audibility of any of these effects is debatable, natch.
HTH - Don
Thanks...i got (some) of that. Really appreciate your taking the time to write that all down for guys like us to read (and re-read) to learn. What is your view about NOS DACs? There is another thread here about NOS DACs where BFlowers, i think, observed his two favorite DACs at CES were both NOS (Ypsilon and Zanden).
Don, thanks for taking the time to write that up, much appreciated!
This is turning out to be a very interesting thread!
This is turning out to be a very interesting thread!
No worries Lloyd, its a blast.
The two that offer amazing bang-for-the-buck are TDA1387 and TDA1545A. I'm currently building up arrays (around 20 chips) using TDA1387 and the first prototype sounds very satiisfying indeed. I don't only like these chips for their sound, they are also very low power (less than 30mW). Their only drawback is they're only 16bits but this is no kind of drawback on redbook material just means to extend my designs to hires I might have to find another chip.
There are very few DACs out in the wild that use these chips. The only one that I know of is rather DIY oriented, not a completed box : http://tech.juaneda.com/en/projects/jundacfive.html. Of course I also like TDA1541A and this is available in several guises as completed units, I just prefer not to design with this chip because its too expensive and power hungry to build into arrays, which are my passion. Besides many gifted designers have embraced the TDA1541A - for myself I like to beat my own path rather than follow the crowd
I'll have a go at generalizing. D-S ( or S-D) don't suffer from glitching and they have very good low level linearity. This means excellent 'bloom' - the soundstage can be very expansive. But they lack what I call 'jump factor' - the dynamic contrasts in the music are reduced, they sound like music has been compressed. DSD (being an implementation of S-D) exhibits this effect - when I listen to DSD recordings of piano, the 'bite' of the piano (piano is after all a percussion instrument) is diminished, the startle of a pianist's sforzando is somehow papered over. To my ears DSD turns music into almost 'muzak' for elevators, it becomes too inoffensive, mellow.
S-D has another effect which is on the tonality of instruments. I first noticed this on massed brass - Tchaikovsky's 4th symphony has long been a favourite of mine, the old Mravinsky with Leningrad Phil is a magical recording. But the blast of brass which opens this has its tonality corrupted on a PCM1792 (S-D DAC) when compared side by side with a NOS DAC.
R2Rs though have different artifacts - their glitchiness means they tend to 'grey out' tonality when used in OS - the faster the clock rate the faster the glitches come. They can also suffer from poorer low-level linearity which means soundstages might not be so deep. To fix up multibit DACs with more oversampling than 2X is just asking for trouble from glitches. NOS though while it has tremendous tonal purity due to lowest glitch density has other drawbacks (freq response droop, imaging above 20kHz) which mean I've abandoned it in my designs. For me, 2X OS is the sweetspot.
Don - one thing missing from your otherwise excellent precis was the observation that whether jitter becomes an audible issue does to some degree depend on the kind of DAC in use. Normal multibit DACs are the best as regards jitter immunity. Due to their heavy out of band noise generation, D-S DACs have the foible that jitter can 'fold down' this noise into the audio band right? Taking your earlier arguments on board that jitter effects scale linearly with frequency and amplitude, the effect of jitter on this high level, high frequency noise ISTM must be significant.
Hmmm... I'll have to think about that. Ideally, the distortion induced by jitter is dependent on the signal frequency, not the clock, so sampling rate does not matter (just resolution and signal frequency -- I did put the derivation in one of those jitter threads, or maybe one of the DAC threads*). Noise shaping moves the sampling (quantization) noise up in frequency; I do not think it aliases (folds) any differently than a normal converter but I have not analyzed it (or rather I have but do not recall it off-hand; most of my designs were conventional segmented DACs).
Note that aliasing is an issue for ADCs: an input signal with bandwidth equal to or greater than 1/2 the sampling frequency will fold back into the signal band. There is no folding at the output of a DAC; the problem is an image of the input signal is created at a frequency "reflected" around 1/2 the sampling frequency. Put a mirror at fs/2 and any signals below are reflected above fs/2. A 15 kHz signal sampled at 20 kS/s produces an image at 25 kHz.
I did not think the image theory was different for a DS DAC but must think about the loop dynamics a bit. I thought since the noise is shaped by the loop, any HF noise or images will be suppressed so they appear at the output "normally" but are reduced by the loop gain back in the input. The net effect is imaging response is the same in a DS DAC as a conventional DAC. I could be wrong...
Disclaimer: With exception of a few specific references, I am discussing DACs in general, not specifically the dCS unit. That means we are somewhat off-topic, for which I apologize. I have obviously never heard the dCS so cannot comment upon its sound and none of my comments should be taken as specific to the dCS. Hopefully I will avoid disparaging anyone, but frankly when engaged in technical discussion I tend to post what I am thinking without thought about the feelings of others. I tend to think we can learn a lot from each other, subjective or objective.
Onwards - Don
* Found it: http://www.whatsbestforum.com/showthread.php?1322-Jitter-101
p.s. Thanks all for the kind words! I really appreciate them, just get focused on the discussion and don't always say so...
p.p.s. There are conventional DAC architectures that reduce glitches significantly, such as segmented or fully-unary designs. See the DAC threads, I am already taking this thread way off-topic.
Note that aliasing is an issue for ADCs: an input signal with bandwidth equal to or greater than 1/2 the sampling frequency will fold back into the signal band. There is no folding at the output of a DAC; the problem is an image of the input signal is created at a frequency "reflected" around 1/2 the sampling frequency. Put a mirror at fs/2 and any signals below are reflected above fs/2. A 15 kHz signal sampled at 20 kS/s produces an image at 25 kHz.
I did not think the image theory was different for a DS DAC but must think about the loop dynamics a bit. I thought since the noise is shaped by the loop, any HF noise or images will be suppressed so they appear at the output "normally" but are reduced by the loop gain back in the input. The net effect is imaging response is the same in a DS DAC as a conventional DAC. I could be wrong...
Disclaimer: With exception of a few specific references, I am discussing DACs in general, not specifically the dCS unit. That means we are somewhat off-topic, for which I apologize. I have obviously never heard the dCS so cannot comment upon its sound and none of my comments should be taken as specific to the dCS. Hopefully I will avoid disparaging anyone, but frankly when engaged in technical discussion I tend to post what I am thinking without thought about the feelings of others. I tend to think we can learn a lot from each other, subjective or objective.
Onwards - Don
* Found it: http://www.whatsbestforum.com/showthread.php?1322-Jitter-101
p.s. Thanks all for the kind words! I really appreciate them, just get focused on the discussion and don't always say so...
p.p.s. There are conventional DAC architectures that reduce glitches significantly, such as segmented or fully-unary designs. See the DAC threads, I am already taking this thread way off-topic.
Yes - but in the case of a D-S DAC, the signal coming out of the DAC is a mixture of the wanted audio band signal (at the left side of the freq axis on an FFT plot) and noise which is the big lump on the right hand side. In the case of 1bit converters (like DSD), the noise is always at a higher level than the signal, just moved a distance away on the frequency axis. Its the effect of jitter on this noise which intrigues me. Taking your requirement of 100pS for full scale at 20kHz to result in <16bit level artifacts, ISTM that (taking DSD) we'd need at least 64X better jitter than this to ensure none of that noise gets sidebands which fall into the audio band. That gives me a requirement of <2pS (DSD clocks at 64X). But in actuality the situation seems worse than this because the noise is always there irrespective of signal level - it does not scale with signal as in the case of the multibit DAC.
That does not match my understanding of the DS transfer functions... The jitter requirement is based upon the aperture time, which in turn is related to signal frequency and resolution. The DS approach would be in serious trouble if the jitter was multipled by the over-sampling ratio (OSR) as you imply, unless I am not understanding your post (quite possible). See texts by Temes, Candy, and Norsworthy.
The high-frequency noise is much higher, true, reaching unity at Nyquist, the trade for suppressing the noise at low frequencies. However, that unity peak is reached at a frequency 64x the signal band (using your example) where it easier to filter. The total quantization noise energy is the same, but instead of being flat across the band, it is greatly reduced at LF and pushed into the HF side, well above the signal band.
I am not sure why the effect of jitter on noise matters? The noise, and HF jitter, is filtered out by the analog output filter (typically of order N+1 for an Nth-order modulator, but higher order is OK since the cutoff frequency can be well above the signal band. In-band noise and HF jitter is reduced by the loop since quantization noise and jitter both occur at the sampling point (and are thus reduced by the loop gain).
I feel I am not following you; I'll try again tomorrow after sleeping on it (too bad we are on opposite sides of the world!) - Don
The high-frequency noise is much higher, true, reaching unity at Nyquist, the trade for suppressing the noise at low frequencies. However, that unity peak is reached at a frequency 64x the signal band (using your example) where it easier to filter. The total quantization noise energy is the same, but instead of being flat across the band, it is greatly reduced at LF and pushed into the HF side, well above the signal band.
I am not sure why the effect of jitter on noise matters? The noise, and HF jitter, is filtered out by the analog output filter (typically of order N+1 for an Nth-order modulator, but higher order is OK since the cutoff frequency can be well above the signal band. In-band noise and HF jitter is reduced by the loop since quantization noise and jitter both occur at the sampling point (and are thus reduced by the loop gain).
I feel I am not following you; I'll try again tomorrow after sleeping on it (too bad we are on opposite sides of the world!) - Don
Yes, you are missing my point here. I'm not saying the OSR ratio multiplies up the jitter - it doesn't. But let's do a thought experiment:
Say we have 100pS of jitter on our clock. We know we can just about reproduce a 20kHz full-scale sinewave without problems at this level of jitter. But what about wanting to reproduce a full scale sinewave at 2MHz? Clearly we need 100X lower jitter or we're going to be subject to sidebands right? Jitter's sidebands depend on the rate of change of the signal and 2MHz has 100X the maximum rate of change than 20kHz.
Yes, that's also my understanding.
Yes it seems likely that we'll be subject to folding down of the noise before the analog filter. I agree once the RF signal is stripped off, it can't do any further harm.
Here's a link to an AES letter where Bob Adams (now at ADI) and Malcolm Hawksford are having a discussion about this effect : http://www.essex.ac.uk/csee/research/audio_lab/malcolmspubdocs/J16 Letter re J15 JAES paper.pdf
This effect is the reason the original 'Bitstream' DACs used switched capacitor technology so they'd be relatively independent of clock jitter. See the datasheet of the TDA1547 for an example.
A 2 MHz signal does indeed need 100x less jitter to provide the same resolution, but that is well outside the audio signal band. The whole premise of the DS approach is to oversample and noise shape so that quantization noise is reduced in the signal band. If you want to capture a 2 MHz signal, you need to move the sampling rate up by the same 100x factor to achieve the same SNR as you had with a 20 kHz input signal (all else being equal). Or do something else, like add bits to the modulator, interleave them (perhaps with a Hadamard sequence approach), or whatever.
The quantization noise and jitter is generated at the conversion point, i.e. at the DAC cell at the output of the modulator, and thus shaped by the loop. It is not "folded" to baseband but treated like anything else in the noise transfer function of the loop. A simplified explanation of the transfer fuctions is provided in http://www.whatsbestforum.com/showthread.php?2487-DACs-102-Delta-Sigma-DACs . There were a lot of arguments about noise and jitter, but empirical results showed it not to be any worse than "normal" DACs. Sort-of; there are caveats at the upper end and variations with loop order that ideal theory (and the usual linearized transfer function model) does not handle well. The text by Norsworthy and Candy describes it very well (can't recall the exact title but it's a classic, essentially an extension of the earlier IEEE tutorial and compilation by Candy and Temes; my copy is not on my shelf so I need to check who borrowed it last!)
There are some tricky design coniderations and trades with continuous-time modulators; those are mainly the type I have designed as they ran too fast (GS/s and up) to use SC circuits. SC circuits do solve some tricky jitter issues using CDS (correlated double sampler) cells and such, but I do not think they change the fundamental impact of aperture time variations nor change the image transfer function. There are pros and cons with either approach (continuous-time or switched-capacitor).
I thought Bob retired some time ago? I never met Dr. Hawksford.
The quantization noise and jitter is generated at the conversion point, i.e. at the DAC cell at the output of the modulator, and thus shaped by the loop. It is not "folded" to baseband but treated like anything else in the noise transfer function of the loop. A simplified explanation of the transfer fuctions is provided in http://www.whatsbestforum.com/showthread.php?2487-DACs-102-Delta-Sigma-DACs . There were a lot of arguments about noise and jitter, but empirical results showed it not to be any worse than "normal" DACs. Sort-of; there are caveats at the upper end and variations with loop order that ideal theory (and the usual linearized transfer function model) does not handle well. The text by Norsworthy and Candy describes it very well (can't recall the exact title but it's a classic, essentially an extension of the earlier IEEE tutorial and compilation by Candy and Temes; my copy is not on my shelf so I need to check who borrowed it last!)
There are some tricky design coniderations and trades with continuous-time modulators; those are mainly the type I have designed as they ran too fast (GS/s and up) to use SC circuits. SC circuits do solve some tricky jitter issues using CDS (correlated double sampler) cells and such, but I do not think they change the fundamental impact of aperture time variations nor change the image transfer function. There are pros and cons with either approach (continuous-time or switched-capacitor).
I thought Bob retired some time ago? I never met Dr. Hawksford.
Indeed, but a DSD DAC is doing something remarkably similar to producing a 2MHz tone all the time. Sure the 'tone' isn't a tone, its chaotic, shaped noise but its all HF and therefore it has a high rate of change. Given its presence we need to pay serious attention to the clock to such a DAC if we don't want that HF noise to spray sidebands all over our nice low audio band noise floor.
I know Cirrus use switched caps (as do AKM who I believe are Cirrus' foundry) and so do Wolfson. I haven't seen though that ADI and BB/TI are using this approach so I wonder what their solution to it is. Of course its reduced by having more than one bit in the DAC so the out of band noise can be at a lower amplitude than the wanted signal. Clearly Bob Adams is fully aware of the issue, given his response to Dr. Hawksford. Seeing as we're on a thread about the Vivaldi, I also wonder how dCS address this issue. It doesn't look like they're using switched capacitors - they have latches and resistors. Enquiring minds would like to know...
Given the nature of the chaotic noise coming out of the modulator, I doubt very much whether conventional measurements are attuned to 'seeing' the results. Its why we need our ears to cross-check. FFTs show spurs in narrow bands but I suspect we need wavelet transforms to get a hold of the short-term changes to the (broadband) noisefloor.
I've not met either personally, but last year Scott Wurcer (on DIYA) said he'd pass on my question to Bob. I didn't get a response though - must've gotten lost in the modulation noise
Similar threads
Members online
- five
- Down in Devon
- Armsan
- The Prof
- Massimo66
- 3rdRock
- Porsche
- kmyl
- andromedaaudio
- Hyperion
- unboxed
- ozzzy
- Redwoodaudio
- 2ndLiner
- dminches
- GSOphile
- tdudnyk
- cdka
- Argonaut
- Geoffkait
- Gene 15352
- MRJAZZ
- MattSPL
- matthias
- Elliot G.
- Sarge
- 7ryder
- Król_Artur
- Folsom
- hopkins
- af428
- ticktock
- bozebuttons
- MarcelNL
- Andreas510
- Sun King
- PYP
- trekpilot
- wokeuptobose
- ketcham
- matsurus
- RCanelas
- Salectric
- AudioAbe
- asdffx
- Bonesy Jonesy
- admin1959
 | Steve Williams Site Founder | Site Owner | Administrator |  | Ron Resnick Site Co-Owner | Administrator |  | Julian (The Fixer) Website Build | Marketing Managersing |