Audio Myth - "DSD Provides a direct stream from A/D to D/A."

[Maldur]

PCM with sample rate 22.5 MHz keep more information than DSD512 (22.5 MHz too)

 

22,5MHz PCM don't exist on real life for audio, DSD512 on other hand is playable even my own home right now.

 

Because DSD not acting by f/2 rule, so not limited to f/2 frequency, we can say that DSD can hold more information than PCM - when we have more information .

[audiventory]

Because DSD not acting by f/2 rule, so not limited to f/2 frequency, we can say that DSD can hold more information than PCM - when we have more information .

 

I prefer simple information estimating formulas of information:

 

Info = [bit depth] x [sample rate]

 

DSD 512: 1 x 44100x512 = 22 579 200

 

PCM 352/64: 64 x 352800 = 22 579 200

 

However, we have limitation for pushing extra noise out of useful band due noise shaping limited by computing power.

 

I.e. PCM in reality carrier more information for same conditions.

 

Also any DSD processing demands more significant power.

[Hiro]

I prefer simple information estimating formulas of information:

 

Info = [bit depth] x [sample rate]

 

DSD 512: 1 x 44100x512 = 22 579 200

 

PCM 352/64: 64 x 352800 = 22 579 200

 

However, we have limitation for pushing extra noise out of useful band due noise shaping limited by computing power.

 

I.e. PCM in reality carrier more information for same conditions.

 

 

In reality 64bit PCM DACs don't exist.

 

The two most recent PCM DACs from Schiit, a rare company that still produces such devices, were 20 and 18-bit.

[Jud]

I prefer simple information estimating formulas of information:

 

Info = [bit depth] x [sample rate]

 

DSD 512: 1 x 44100x512 = 22 579 200

 

PCM 352/64: 64 x 352800 = 22 579 200

 

However, we have limitation for pushing extra noise out of useful band due noise shaping limited by computing power.

 

I.e. PCM in reality carrier more information for same conditions.

 

Also any DSD processing demands more significant power.

 

Yuri, then why/how did SDM/PDM wind up being less expensive for A/D and D/A? Less expensive hardware for filtering, I guess, but why less expensive if more power needed?

[audiventory]

What is the difference between "multi-bit DSD" and PCM?

 

We can either increase DSD sample rate or increase DSD bit depth. Transmitted information will same

 

More wide band of DSD need as place for noise, pushed out audible band, only.

[mansr]

Yuri, then why/how did SDM/PDM wind up being less expensive for A/D and D/A? Less expensive hardware for filtering, I guess, but why less expensive if more power needed?

 

A dedicated sigma-delta modulator doesn't need much hardware. Doing it with a general-purpose computer or even a generic DSP is different.

[Miska]

What is the difference between "multi-bit DSD" and PCM?

 

Multi-bit SDM is just like DSD but with multiple quantization levels. And actually DSD can be also 1-bit and multi-bit at the same time...

 

Generally with SDM, bandwidth available for the signal is not clearly defined and sample values don't encode instantaneous signal values. Characteristic property is existence of sigma-delta modulator in the chain. It becomes more clear closer to the A/D or D/A conversion stage you get. Multi-bit SDM DAC operates on scrambled unary coded values (aka thermometer code) while multi-bit PCM DAC operates on binary coded integer values.

 

But there are also lot of representations of signal that is neither PCM nor SDM. I would also categorize floating point representations into this category, because AFAIK it is not usable by any real world converter schematic without requantization to binary encoded integers.

[Jud]

A dedicated sigma-delta modulator doesn't need much hardware. Doing it with a general-purpose computer or even a generic DSP is different.

 

This doesn't quite get to my "bottom line," though. Why doesn't it need much hardware if it needs significant power, or conversely, why wouldn't PCM require even less?

[Miska]

A dedicated sigma-delta modulator doesn't need much hardware. Doing it with a general-purpose computer or even a generic DSP is different.

 

A proper one does take quite a bit. For example Charles Hansen (of Ayre) said that the modulator they have in their A/D converter just barely fits into the FPGA, while decimation filter used to produce PCM output takes only small part of the chip. And for that reason they need to reprogram the FPGA when switching between DSD and PCM outputs.

 

A good modulator also requires really high clock speed which is not usually available in these converter chips..

 

With my algorithms the load is about 50/50 between oversampling digital filters and the modulator.

[audiventory]

Yuri, then why/how did SDM/PDM wind up being less expensive for A/D and D/A? Less expensive hardware for filtering, I guess, but why less expensive if more power needed?

 

Hi Jud,

 

DSD for end user (music lover) is less expensive. In conditions huge available HDD, even for mobile players.

 

As said before, DSD DAC is filter + 2 reference voltage instead PCM's filter + several voltages and/or sigma delta modulator.

 

As far I know, DSD ADC not simpler than currently used PCM (non voltage matrix).

 

 

 

Full music workflow:

 

ADC - PROCESSING - DAC

More power calculations need for processing stage. Also need more power (expensive) hardware/computers.

 

Engineer decissions in processing area are more expensive.

 

Here each optimization of quality/performance ratio is subject of researching (new knowledges practical applying domain). It is also time x money.

 

I.e., theoretically, DSD music production stage is more expensive than PCM. Especially if apply most accurate DSD processing.

 

Either faster and expensive (development and/or hardware), or slow and cheaper, or faster and quality worse (algorithm simplification).

[Miska]

Yuri, then why/how did SDM/PDM wind up being less expensive for A/D and D/A? Less expensive hardware for filtering, I guess, but why less expensive if more power needed?

 

I'm not Yuri, but in short SDM is less expensive and better performing for A/D and D/A hardware because of component tolerances and settling time requirements. Laser-trimming of on-chip resistors for PCM ladder is expensive and slow and you quickly hit sampling rate limits due to settling time / accuracty requirements, requiring more complex analog filters which in turn reduce performance...

 

Although SDM takes much more DSP power, it is relatively cheaper. Advances in available computing power made SDM converters feasible, before that PCM ladders were the only technically feasible way. There's still lot of space for improvement in SDM area using more DSP power. Hardware chip implementations have got stuck to what I'd call "2nd generation".

[mansr]

A proper one does take quite a bit. For example Charles Hansen (of Ayre) said that the modulator they have in their A/D converter just barely fits into the FPGA, while decimation filter used to produce PCM output takes only small part of the chip. And for that reason they need to reprogram the FPGA when switching between DSD and PCM outputs.

 

That's still a sliver of a fraction of the size of a modern CPU. If realised in silicon, it could easily be done in the cheapest process available with room to spare.

 

A good modulator also requires really high clock speed which is not usually available in these converter chips..

 

With my algorithms the load is about 50/50 between oversampling digital filters and the modulator.

 

A simple SDM (i.e. no fancy features like lookahead) with an 8th order filter can achieve perfectly acceptable performance. A dedicated hardware implementation could in principle be constructed to run directly off the bitstream clock. Since the clock rates are anyhow quite low, it probably makes practical sense to use fewer computational blocks and run at something like 8x the bit clock.

 

Specialised hardware is almost always a few orders of magnitude more efficient at doing its one task than is a general purpose computer.

[Jud]

Yuri and Miska, thanks, the multiple reference voltages/resistor trimming for PCM hardware vs. the higher computing power for SDM (but still cheaper overall than PCM) is what I was looking for.

[mansr]

Multi-bit SDM is just like DSD but with multiple quantization levels. And actually DSD can be also 1-bit and multi-bit at the same time...

 

Generally with SDM, bandwidth available for the signal is not clearly defined and sample values don't encode instantaneous signal values. Characteristic property is existence of sigma-delta modulator in the chain. It becomes more clear closer to the A/D or D/A conversion stage you get. Multi-bit SDM DAC operates on scrambled unary coded values (aka thermometer code) while multi-bit PCM DAC operates on binary coded integer values.

 

If you take two DSD bitstreams and add them bit for bit, you end up with something where each sample can take on values [0,2]. Is this 2-bit DSD or 2-bit PCM? Does it matter?

[audiventory]

A simple SDM (i.e. no fancy features like lookahead) with an 8th order filter can achieve perfectly acceptable performance.

 

Happyness not in order number, but in skill of viewing new possibilities in given order

[audiventory]

Yuri and Miska, thanks, the multiple reference voltages/resistor trimming for PCM hardware vs. the higher computing power for SDM (but still cheaper overall than PCM) is what I was looking for.

 

Jud, I'm glad, that been able give you useful information

[Jud]

Jud, I'm glad, that been able give you useful information

 

Even better, you and Miska gave me useful software.

 

Edit: Well not exactly "gave," I bought it.

[mansr]

Even better, you and Miska gave me useful software.

 

Edit: Well not exactly "gave," I bought it.

 

Mine is free :-)

[Jud]

Mine is free :-)

 

And, one would hope, more than worth the price.

[audiventory]

Even better, you and Miska gave me useful software.

 

Edit: Well not exactly "gave," I bought it.

 

Thank you for kind words. It inspires me for further work!

[mansr]

And, one would hope, more than worth the price.

 

It works for me. If someone else finds it useful, all the better.

[Jud]

It works for me. If someone else finds it useful, all the better.

 

Link and short description?

[tailspn]

Another consideration between PCM and DSD, and the reason for the much lower PCM sampling rate verses DSD bit rate is the nature's of PCM and DSD.

 

PCM is made of individual binary integer words that stand alone, like the frames of motion picture film. In motion picture film each frame is an individual complete picture. Sequenced together, they form motion. But the amount of redundant data from one frame to the next is a very high percentage. The same is true with PCM samples representing a analog level in a 2's complement binary word each sample time. From sample to sample, not only is the change data is represented, but by the nature of the format, the constant data is dragged along. That's the primary reason every other data transmission discipline abandoned PCM years ago; lack of efficiency.

 

DSD has a similar problem; the amount of data describing the out of band of interest uncorrelated shifted noise. But at least over any segment of time, only the change information is represented, allowing much high rates. As higher bit rates are employed, with the band of interest remaining constant, the gap between the band of interest and beginning levels of uncorrelated shifted noise expands, allowing for even gentler more phase linear filtering to occur. It's the filtering that affects the sound quality of music we listen to, not the storage format, and the necessity to employ filters when converting formats.

 

PCM is an archaic tricycle with training wheels dying format. It persists in audio music production because, with its minute market size, there isn't the necessary capital investment for the development of DSD production tools.

[mansr]

Link and short description?

 

I've made contributions to many open source projects. Of greatest relevance to this thread is probably the DSD encoder I recently added to SoX. There's another thread on here with links and some graphs.

[tailspn]

If you take two DSD bitstreams and add them bit for bit, you end up with something where each sample can take on values [0,2]. Is this 2-bit DSD or 2-bit PCM? Does it matter?

 

Only in the sense that PCM isn't a generic term, it's an accepted encoding practice with specification. To apply it to a different encoding principle may be very confusing to some. Like me.