Debate of DAC design regarding DSD vs PCM among 5 VIPs

[rredline]

1 hour ago, semente said:

 

Watts talk s a bit about taps here:

 

http://www.the-ear.net/how-to/rob-watts-chord-mojo-tech

 

“Mojo has 500 times more processing power than conventional high performance DACs”

1 hour ago, semente said:

 

R

 

 

Sorry I think this is BS

To achieve ASIC chip level performance, you need a much more complex and power hungry FPGA, certainly not that kind of entry level FPGA in Mojo capable of.

[yamamoto2002]

4 hours ago, rredline said:

 

 

Disagree

If DSD recording causes non-linear time error, PCM will too, because that PCM is came from DSD after decimation filter applied.

 

I think he'd pointed out about timing error caused by noise shaping function. Rounding error on sample #n will be carried over to the sample value #n+1, but it should represent the sample value of 1 sample time later (original analog signal information on time #n partly moved to time #n+1 and the quantity moved is varied by rounding error value from quantization of time #n) and this causes timing error. I'm not sure if the phenomenon of this explanation is truly problematic or not

[louisxiawei]

On 4/12/2017 at 4:11 AM, Nikhil said:

 

Thanks for adding this.  I have also wondered what the term "taps" meant.
I don't think I still understand it but it's a starting point. Need to read up on this some more.

 

 

I'm not sure if this is a translation issue but I think that the correct word in English would be "number" instead of "length".  Chord always talks about the number of taps increasing the accuracy etc.  audio.bill's post also refers to number of (FIR) taps etc.  Can someone confirm this?

 

Regards.  

 

Just have a look at the technical specifications of Hugo 2 by glance.

 

It shows "Tap-length" : 49152. So right now I'm not sure whether its a tap length or tap number. In the original article, it is "tap-length" indeed from direct translation.

 

Guess we need someone's further confirmation.

 

[audiventory]

10 hours ago, semente said:

I think he probably means forms of distortion that affect the recreation of the spatial effects you find in a recording, e.g. higher noise floor will reduce or mask the decay of instruments.

 

Hi Semente,

Yes. Analog noise floor really have the masking feature.

Also we can consider smooth changing of linearity of tape, that depend on level.

[audiventory]

5 hours ago, yamamoto2002 said:

I think he'd pointed out about timing error caused by noise shaping function. Rounding error on sample #n will be carried over to the sample value #n+1, but it should represent the sample value of 1 sample time later (original analog signal information on time #n partly moved to time #n+1 and the quantity moved is varied by rounding error value from quantization of time #n) and this causes timing error. I'm not sure if the phenomenon of this explanation is truly problematic or not

 

There is too little information for understanding, that he meant as DSD timing error. Need take scheme and discuss it.

[louisxiawei]

Interview with Ted Smith from PS audio

 

4.1 Before doing the D-A conversion on DirectStream DAC, no matter PCM or DSD signal, they are both being converted to the DSD first. Are you suggesting that DSD is superior to PCM?

 

Before discussing it, I must give some guide of the basic DSD concept so that normal people can understand it better.

 

In terms of the format, DSD is equally good as PCM. They both can record a huge amount of music information. The difference is just the format. The reason I insist on DSD development is not because PCM this format is not good enough, but rather the way of PCM D-A conversion is not ideal enough. Simply put: multi-bit PCM dac’s structure is too complex and it can only rely on digital filer to “predict” the original analogue waveform, which can somehow mask the original “appearance” of the digital signal. In other words, PCM dac cannot exploit the advantage of PCM format’s information to the full.

 

But for the DSD DAC, it is such an ideal D-A conversion approach for me to be much simpler, more linear and closer to the analogue. In theory, DSD dac only needs one resistor and one capacitor to form a simple low-pass filter circuit, which converts DSD signal to analogue signal. This kind of conversion not only has much lower error, but also doesn’t need strict component matching like PCM dac. However, please note that these advantages I mentioned above can be only achieved by 1 bit DSD D-A conversion. On that premise, I personally think 1 bit DSD dac is much better than multi-bit PCM dac even for processing the PCM signal. If PCM signal is converted to the DSD first, 1 bit DSD dac will be capable of exploiting all hidden information in that PCM-converted DSD signal.

 

In conclusion, no matter PCM format or DSD format, 1 bit DSD dac is the most ideal approach for D-A conversion.

[jabbr]

On 4/14/2017 at 4:45 AM, rredline said:

 

 

Sorry I think this is BS

To achieve ASIC chip level performance, you need a much more complex and power hungry FPGA, certainly not that kind of entry level FPGA in Mojo capable of.

You'd be surprised what "entry level" FPGAs can do nowadays. I think the FPGA that's used in these DACs is about $20 and uses 1w or so... but there is considerable IP needed to optimize these chips whether ASIC or FPGA. Nowadays a "more complex and power hungry FPGA" might have a couple of onboard 4 Ghz ADC/DAC, be used to implement 100g Ethernet or build a digital oscilloscope/FFT analyzer, smart vision system for autos etc etc. 

[Superdad]

On 4/14/2017 at 5:41 AM, louisxiawei said:

Just have a look at the technical specifications of Hugo 2 by glance.

 

It shows "Tap-length" : 49152. So right now I'm not sure whether its a tap length or tap number. In the original article, it is "tap-length" indeed from direct translation.

 

Each "tap" is nothing but a pair of filter coefficients.  Number of taps = number of coefficient pairs.  The greater the number of taps, the "longer" the filter is.

 

So "tap-length" is a misnomer for filter length--which is just number of taps.  Hope that is more clear. 

[JohnSwenson]

5 hours ago, Superdad said:

 

Each "tap" is nothing but a pair of filter coefficients.  Number of taps = number of coefficient pairs.  The greater the number of taps, the "longer" the filter is.

 

So "tap-length" is a misnomer for filter length--which is just number of taps.  Hope that is more clear. 

Almost,

an FIR filter is a sequence of delay lines, the connection between each delay element is "tapped off", run through a multiplier (the "tap coefficient") and and the result of these multiplies are all added together. The value before the first delay element is also tapped.

 

 

The filter length is the number of delay elements, the number of taps is the filter length plus 1.

 

In digital FIRs the delay elements are flip flops, arranged as registers, for audio data something like 32 bits longs. Each delay element is one of these registers. The output of a register feeds into the next, all clocked by the same clock. The hard part is the multiplier. Each tap takes its value and multiplies its coefficient. (usually considered to be between zero and one), this multiply takes significant hardware to implement.

 

Thus for a 100 length filter you have 100 32 bit registers, 101 taps, each of which is a 32x32 bit multiplier, and one 101 input adder, it adds all the 101 numbers from the multipliers together. The output of the adder is the "output" of the filter for that particular time slice.

 

The value of those 101 coefficients determines what the filter does.

 

Even a small number of taps can do significant filtering. The more taps you have the more control you have over exactly what the filter does. A large number of taps does not necessarily produce a "better" filter, it gives the designer more control to implement exactly the filter they want.  

 

John S.

 

 

[Miska]

11 hours ago, JohnSwenson said:

Even a small number of taps can do significant filtering. The more taps you have the more control you have over exactly what the filter does. A large number of taps does not necessarily produce a "better" filter, it gives the designer more control to implement exactly the filter they want. 

 

Yes, the number of taps is not useful as measure of filter quality, it is pretty much arbitrary number.

 

It is not a problem run a million tap filters on computer. For example room correction filters are typically 65000 to 256000 taps and processing that takes negligible amount of CPU time at PCM rates, including 768k... When you have something like 16 million taps at 24 MHz sampling rate, then we start talking about noticeable amount of CPU time.

[louisxiawei]

Interview with Ted Smith from PS audio

4.3 But many people think DSD dac has high-frequency problems, what's your opinion?

 

The high-frequency problem does exist, therefore DSD dac needs Noise shaping to solve it. By taking advantage of the Noise Shaping, the very high-frequency noise can be eliminated and keep the intact information of the music, which our ears can hear.

 

To solve this problem completely, DirectStream DAC does a 10X upsampling for all DSD signal beofore D-A conversion so as to push the noise to the very high-frequency area that human ears cannot detect, then covert back to 2x DSD. I think this can completely solve the so-called “inborn flaw” of the DSD dac.

[Gordian]

Thanks Louixiawei for this great topic

[wendysire523]

@louisxiawei

Hi Louis, thanks for sharing. Im interested in 5 VIPs interviews but the magzine in the link you shared not free for this part. Really appreciate if you could help scan and post those pages with higher resolution again! Or in case you prefer email: [email protected] 

 

Thanks very much.

wendy

[bibo01]

@louisxiawei

I am interested in all Andreas Kock interview.

Take your time.

Thanks 

[louisxiawei]

Interview with Ted Smith from PS audio

4.6 Don't you think the conversion between DSD and PCM will cause any distortion?

 

The conversion between DSD and PCM is used by FPGA I developed for nearly 10 years, which can make sure that the distortion from the conversion to the lowest. Many would think it might be better with less conversion, but the merit of this conversion is greater than its disadvantage. As I mentioned, in terms of format, DSD is equally good as PCM, the key is to use 1 bit DSD DAC to do the D-A conversion.

 

 

I personally am a little disappointed regarding Ted's 4.6's answer. It means nothing to me. I hope someone can give us more detailed answers about the potential lossy signal/distortion from the DSD-PCM conversion. 

 

 

@Em2016 All your interested parts have been translated.  

 

@bibo01 I will focus on Andreas' bits now. 

 

[asdf1000]

3 minutes ago, louisxiawei said:

Interview with Ted Smith from PS audio

4.6 Don't you think the conversion between DSD and PCM will cause any distortion?

 

The conversion between DSD and PCM is used by FPGA I developed for nearly 10 years, which can make sure that the distortion from the conversion to the lowest. Many would think it might be better with less conversion, but the merit of this conversion is greater than its disadvantage. As I mentioned, in terms of format, DSD is equally good as PCM, the key is to use 1 bit DSD DAC to do the D-A conversion.

 

 

I personally am a little disappointed regarding Ted's 4.6's answer. It means nothing to me. I hope someone can give us more detailed answers about the potential lossy signal/distortion from the DSD-PCM conversion. 

 

 

@Em2016 All your interested parts have been translated.  

 

@bibo01 I will focus on Andreas' bits now. 

 

 

Thank you so much !! Your efforts are greatly appreciated.

[louisxiawei]

Interview with Andreas Koch from playback

2.1 Your playback DAC convert files no matter PCM or DSD into DSD first and then do the D-A conversion afterwards, do you think it is a better way to do this even for the PCM signal?

 

The multi-bit PCM DAC has its non-linear distortion problem because the proportion of each bit’s output is different. To decrease this non-linear distortion, the dac has to reply on high-precision components and circuit, and consequently, leading to high manufacturing cost.

 

Moreover, most of multi-bit PCM dacs use Brickwall filters that can produce pre-ringing. Generally speaking, this pre-ringing is the sound that most people loathe and being described as “ annoying digital sounding” quite often. By comparison, DSD dac is much more linear and do not need digital filter, which can solve many PCM DACs’ flaw.

[louisxiawei]

Interview with Andreas K

2.2 Many people think only 1 bit DSD D-A conversion is the genuine and correct way to work, what do you think?

 

Yes, that’s true. The premise of DSD dacs being better than PCM ones is that DSD dacs must be a true 1-bit DSD modulation. However, the problem is that most of Delta sigma dac chips on the market nowadays will make the DSD signal get through low-pass-filter circuit first to be converted to lower-sampling multi-bit signal before doing the D-A conversion. Under that circumstance, DSD signal in fact has been converted to PCM one already, which can also cause many PCM-related problems. As far as I know, most of DAC chips on the market are in the category of multi-bit Delta sigma dac modulation, which all convert DSD signal to 5-6 bit PCM signal.

 

Only by building separate circuit for DSD and PCM, we then can achieve 1 bit native DSD D-A conversion, and all Playback Designs dacs are made in this fashion.

[semente]

7 hours ago, louisxiawei said:

Interview with Andreas K

2.2 Many people think only 1 bit DSD D-A conversion is the genuine and correct way to work, what do you think?

 

Yes, that’s true. The premise of DSD dacs being better than PCM ones is that DSD dacs must be a true 1-bit DSD modulation. However, the problem is that most of Delta sigma dac chips on the market nowadays will make the DSD signal get through low-pass-filter circuit first to be converted to lower-sampling multi-bit signal before doing the D-A conversion. Under that circumstance, DSD signal in fact has been converted to PCM one already, which can also cause many PCM-related problems. As far as I know, most of DAC chips on the market are in the category of multi-bit Delta sigma dac modulation, which all convert DSD signal to 5-6 bit PCM signal.

 

Only by building separate circuit for DSD and PCM, we then can achieve 1 bit native DSD D-A conversion, and all Playback Designs dacs are made in this fashion.

 

This is interesting...

I wonder what 1-bit SDM chips he uses.

 

R

[rredline]

8 hours ago, louisxiawei said:

Interview with Andreas K

........most of DAC chips on the market are in the category of multi-bit Delta sigma dac modulation, which all convert DSD signal to 5-6 bit PCM signal

 

 

Disagree.

 

It is the flaw of 1-bit SDM engine that caused DAC chip maker to produce 5-6bit SDM engine. DSD is not converted to PCM in these engines as there is no decimation fllter applied in the process.

 

Mr Andreas has been saying the SAME thing again and again in various interviews/workshops, It is a bit disappointing that a man with his level of expertise, so many of tasteless marketing talks and so little of technical tidbits have been revealed to the community. 

[Superdad]

5 hours ago, rredline said:

 

 

Disagree.

 

It is the flaw of 1-bit SDM engine that caused DAC chip maker to produce 5-6bit SDM engine. DSD is not converted to PCM in these engines as there is no decimation fllter applied in the process.

 

I too was shocked to read that mis-information from Mr. Koch.  At first I assumed that the translation was in error, but to see you report that he has made such a false claim elsewhere leaves me scratching my head.

 

Multi-level PDM (as used in S-D DAC chips and in shift-register discrete designs such as Miska's DSC-1 and others) is in no way shape or form similar to binary-twos-complement multibit PCM!  Not in rate, not in encoding, not in resolution, and not in where it puts noise.

 

Remember folks, DSD is just the marketing name for 1-bit (well really two-level) variant of PDM (pulse data modulation), and if done right, there are lots of good reasons for a designer to harmlessly move a 1-bit DSD/PDM stream into a multi-level PDM format.

 

[That said, the very different, and quite critical steps of taking Redbook and interpolating that first to high rate PCM and then through SDM modulators to produce a PDM stream--that's where the art comes in, ala HQ Player or some of the bettter DACs with the horsepower and carefully refined code/filters for the task.

By the way, if we had a home computer interface and cable protocol to send multi-level PDM (call it DSD-wide if you wish), then DACs could be built for that and Jussi would likely be happy to oblige multi-level output from his fine SDM engine.]

[Miska]

8 hours ago, semente said:

This is interesting...

I wonder what 1-bit SDM chips he uses.

 

PBD doesn't, AFAIK, use DAC chips. It is all built around FPGA...

 

[Miska]

There is some strange misconception that multi-bit SDM DACs would be related to PCM and operate in similar fashion. It would be plain stupid to do it like that. PCM is wasting a lot of bits. If we think a sinewave that is -12 dB level, it means that top two bits are never used to represent it, that already loses a huge range of values, only quarter is used. In addition, many more of the bits are only used around peak regions of the waveform. Only good thing is that this reduces also biggest problem of R2R ladders because there the biggest errors come from MSBs. But there is very little resolution left at the lowest levels / around zero-crossing.

 

For example TI/BB chips use 5-level SDM which would be about 2.3 bits in PCM terms. Meaning that anything below -6 dBFS level would be 1-bit anyway and the MSB would practically always unused. Even for "6-bit" in PCM style, signal would become 1-bit below -36 dB. However, SDM always uses all bits/levels/elements, regardless if signal level. It is most precise at the lowest levels / around zero-crossing.

 

My DSC1 DSD-DAC is "5-bit" if you think in terms of two's complement -> log2(32) = 5, or more precisely log2(33) = 5.0444 bits. Although that kind of bit figure is completely useless in this context. The meaningful part is that it can produce 33 different output current levels which are turned into voltage levels by I/V stage.

 

[Ralf11]

The Chord guy has really trashed DSD.

 

Then there is that listening study in JAES from 2007.

 

 

[Nikhil]

8 hours ago, Ralf11 said:

The Chord guy has really trashed DSD.

 

Then there is that listening study in JAES from 2007.

 

 

 

The problem I have is that whatever people say negatively about DSD, it doesn't bear with what I hear in my setup.  

If you take the trouble to make the investment in the computing power needed to process DSD (and a DSD DAC), it is a very enjoyable sound.