PCM to DSD (DSC1) vs native PCM

[Miska]

8 minutes ago, numlog said:

My understanding is fairly limited but AFAIK converting PCM to DSD is similar to how a Delta Sigma DAC works.

How do they differ? Why would one be superior to the other?

 

From the DAC point of view, not so much, for example the Sabre 9038 has 64 unity-weighted elements. DSC1 has 32. If you look at other discrete implementations like dCS RingDACs, they have 24. Chord varies from 4 to something like 20. The rest is about how the element array is used and what kind of analog reconstruction filter there is at the output.

 

Large portion of the performance is defined by the DSP algorithms generating the bitstream. The DAC itself is fairly simple. That is one of the purposes of SDM DACs, that as much as possible things that require accuracy and complexity are done in digital domain using DSP, rather than in analog domain on hardware. This improves reliability and precision while also being cheaper at the same time.

 

11 minutes ago, numlog said:

At a given DSD rate, DSD to PCM conversion can be done very with little computing power or it can require more computing power than what our fasted consumer CPUs are capable of.

what is the difference with these conversions?

 

Complexity of the algorithms. Things like to which rate the digital filters are run to and how the modulators operate. These algorithms can be quite different.

 

13 minutes ago, numlog said:

If you aren't using a PC the option for PCM to DSD conversion would be limited to the AK4137 SRC.

I am interested in the DSC1 for DSD playback regardless but I could pick up an AK4137 board for PCM conversion...

The question is if there would be any point if I already own a decent 9038Q2M DAC when the aim is to achieve the best PCM playback possible.

 

With those chip-converters, you again have similar limitations as you have with the built-in converters of DAC chips like 9038 and the exercise doesn't have much point.

 

Modern DAC chips are mostly designed to a price point, so they are generally below 20EUR/piece. And when necessary, cut some corners to achieve that.

 

[Miska]

1 hour ago, numlog said:

only things start to get confusing when using HQPlayer's best filters werent able to outperform the DACs internal DSP.

 

Which way?

 

On my 9038Q2M based DAC (Pro-Ject PreBox S2 Digital), I have confirmed results of improved digital filter performance and some other aspects. You can find the results from the relevant thread here.

 

1 hour ago, numlog said:

I have wondered how the 9038 DACs handles the DSD input before the SDM portion, for example it does allow volume control.

 

It goes through it's DSP with volume control and remodulation. So it doesn't help working around it's modulator. But it helps working around shortcomings of it's digital filters. Or put other way, DSD goes through much simpler DSP chain in the chip than PCM does.

 

1 hour ago, numlog said:

AKM DAC's have some sort of bypass mode for DSD that bypass volume control, filtering and whatever else before SDM stage, clearly there must be some disadvantage to what Sabre is doing for DSD if AKM has this mode, so the DSC1 might be the only right choice to hear HQPlayer to its full potential

 

Yes, down side of Sabre is that it doesn't have Direct DSD mode. While AKM, Cirrus Logic and Wolfson have. And TI/BB chips don't even have anything else.

 

You can find some DSC1-style DACs on the market that have a separate discrete DSD section. First and most well known is T+A with DAC8 DSD and the newer more expensive models. I guess Denafrips is closest to DSC1. Then there are Holo Audio DACs which I believe are also similar and (AFAIK) also feature AKM SRC chip you mentioned (just like TEAC UD-50x series).

 

Thread here discusses some of these aspects (but also about R2R PCM implementations):

At the end you can find fairly up to date list of DACs.

 

[Miska]

10 minutes ago, PAR said:

Yet I still don't like it compared to unvarnished PCM when an appropriate digital filter has been selected as dCS equipment allows.

 

dCS is never "unvarnished PCM" because it is always an SDM DAC by construction. So it goes through modulators always, just different modulators. It technically cannot play PCM without conversion.

 

With DSD upsampling the choice of digital filter is as important as for any other case too.

 

One limitation of dCS' DSD-upsampling is that it produces usually DSD64 (2.8 MHz) or at most DSD128 (5.6 MHz). These are fairly low rates these days. While I can play up to DSD1024 (49.2 MHz)...

[Miska]

6 hours ago, PAR said:

 Absolutely correct.  But in my view, so what? Do ever higher rates make the music any better ? They may improve the sound marginally but that isn't my question. 

 

I know that is an unfair question for you as you have what Churchill called " a dog in the fight" given your business profile. You have to say " Yes" of course 😉.

 

Yes, in my opinion step from DSD64 -> DSD128 is huge, from DSD128->DSD256 still big, and then it starts getting smaller.

 

Have you tried yourself DSD64 -> DSD128 -> DSD256 -> DSD512? If you haven't, you cannot say "No" to your own question either.

 

And not the rates alone, also the modulator choice matters. So it is good to listen to at least couple of modulator alternatives.

 

[Miska]

On 11/21/2019 at 4:44 PM, numlog said:

What is the resulting difference (subjective and objective) between having 32 elements and 64?

 

Mainly roll-off of the resulting analog FIR filter. My 32-element DSC1 design works OK starting from DSD128 (not too much roll-off at 20 kHz yet). If you intend to run higher rates, then you can increase number of elements as well.

 

On 11/21/2019 at 4:44 PM, numlog said:

Alternatively what about having far less, 4 or 8 elements? I am weary of quality of chinese kit so also considering building a simpler version from scratch... it would cost basically nothing so no reason not to.

 

I would say that is not enough with unity weighting, but you could cut it down to 16- or 24-elements if you like, especially if you plan to run it at DSD64. But you really should look it as entire DAC including the analog reconstruction filter following it, which in my case is 4th order. IOW, what matters is combined performance of the analog FIR DAC + analog reconstruction filter.

 

Also the D/A conversion stage design affects jitter sensitivity.

 

[Miska]

10 hours ago, numlog said:

Unless im interpreting it wrong it sounds like more elements = slower roll off?

I was assuming it would create a steeper filter, but also suspecting it would influence THD and noise performance, so if it doesnt improve THD/noise or make the filter more effective what's the purpose?

 

More elements mean earlier roll off and more attenuation. You double the number of elements and the roll off starts at half the frequency. The roll off point is relative to sampling rate, so the absolute frequency depends on the sampling rate.

 

[Miska]

48 minutes ago, luisma said:

I would have thought this DS DAC would perform / output better sound when fed by DSD vs PCM, your comment of much simpler DSP chain means "better" or "worse" with DSD? I understand it as better but @numlog

 

 

Yes, better... But not sure about the context here...

 

48 minutes ago, luisma said:

@Miska by the way on your Project S2 which filter do you use? fast roll off ?

 

If you send there 705.6/768k PCM, the filter selection has no effect since the built-in filters get bypassed.

 

[Miska]

21 minutes ago, luisma said:

Summarizing, if using HQPe send DSD to the DAC, not PCM.

 

Are the DAC's FIR filters always bypassed when sending DSD?

 

Do you have a specific DAC in mind?

 

[Miska]

Just now, luisma said:

I'm sorry it wasn't clear before, just talking about the Project S2 DAC which you mentioned above

 

Yes, in that case there is no digital oversampling FIR in play when you input DSD.

 

Maybe the confusion is about digital oversampling FIR vs analog D/A conversion FIR? These two are completely different things. For example my DSC1 design discussed here has analog FIR D/A conversion stage, which is fundamental thing for it's operation.

 

[Miska]

On 11/16/2020 at 10:53 PM, Andy-HandyHifi said:

That being, when a DSD signal is present, the ONLY way it is processed is the direct method.  I am positive this is exactly what you were saying. 

 

Yes, indeed...

 

On 11/16/2020 at 10:53 PM, Andy-HandyHifi said:

giving you the the 32 values one would expect of  two's complement 5 bit PCM.  (which of course becomes MUCH MUCH more efficient a way to represent the values once we give to millions of levels in 24 bit and greater ;)

 

Note that switching over to two's complement, while seemingly straightforward, you would loose some of the important presentation power of SDM. Where for most intermediate values you have many possible bit combinations to represent the same value. For example in 4-bit scrambled thermometer code you can represent value "1" in four different ways: 0001, 0010, 0100 and 1000

 

On 11/16/2020 at 10:53 PM, Andy-HandyHifi said:

If we thought of it as we commonly do in 2's complement PCM with negative values, we get this... (assuming we have negatives in unary code... we dont)

 

Most DACs are current output, like DSC1 too, so they don't really have negative values. Just current from 0 to some other value. A chip may have current to voltage converter built-in though. Some chips have this as optional extra, on choice of designer.

 

On 11/16/2020 at 10:53 PM, Andy-HandyHifi said:

My only minor disagreement in what I have independently discovered is that the Signalyst DAC is a 32 level DSM.  When it comes to DSD, it is just a 32 tap FIR filter.

 

No, it is 33-level... ;) With one bit you get two levels, with two bits you get three levels, etc...

 

The register always holds 32 bits, which when latched to output represent once of the 33 possible output levels. The converter itself doesn't understand more. This is always the case, so to get the converter correctly running needs 32 clocks before output is unmuted or otherwise you get loud pop because of uninitialized bits. But this applies to all DSD converters, because you need some time to establish the initial state (some DACs fail to do this and as result you get pops or clicks).

 

On 11/16/2020 at 10:53 PM, Andy-HandyHifi said:

Thank you Miska.  Although the FPGA in my iFi iDSD Pro sounds exceptional, the Signalyst simply by is nature is more capable and I will be using it from now on with Roon/HQPlayer and iDSD direct DSD decode at DSD 1024.

 

Thank you! Great to hear! Modulator design is pretty time consuming and complex task, while it has big impact on the overall performance.

 

And making more and more advanced modulators also make them exceptionally heavy to process. Doing lot of calculations at very high sampling rates means high loads. ASDM7EC being an example of the heaviest so far.

 

On 11/16/2020 at 10:53 PM, Andy-HandyHifi said:

My agreement is why the heck don't we do this all the time??? Why can't be have SDM modulators (other than the Grimm 1 bit converter, discounted here for obvious reasons) output their 4, 8, 16, 32, 64 or 128 level modulator output as a standardised file to be edited, mixed, etc, and THEN downsampled or decimated to the appropriate final format?

 

I guess it is mostly due to convenience factor and legacy. In the past, not very long ago, dealing with something like 96 kHz 24-bit content was heavy and storage consuming enough. So the ADC chips just run the SDM data through PCM conversion, which is also one reason we need to do some fixing with DSP at playback time as well.

 

Most ADC converters use very few levels. If you look at noise floor profiles from ADCs running at 192k or higher PCM output you may notice it looks almost exactly like DSD128 converted to PCM. Because many of the converters run the conversion stage at 5.6/6.1 MHz and only few levels.

 

OTOH, you have multiple DSD128 recorders available (from Korg and TASCAM for example) that use TI's PCM4202 or PCM4204 chip. Which is true 1-bit ADC. Sure it has PCM output mode too, where the SDM stream is converted to PCM as in other chips too. My RME HDSPe AIO cards have PCM4202 ADC (like I believe bunch of other RME converters), but unfortunately lack capability to output the raw DSD data. So you can find many "PCM recordings" on the market, made for example with RME converters, where the data began it's life as real "DSD" and got converted to PCM by the on-chip conversion.

 

In addition, I'm pretty happy with my two RME ADI-2 Pro's and their DSD256 ADC output. Although alternatively you can get very powerful 768/32 PCM output as well.

 

[Miska]

59 minutes ago, Andy-HandyHifi said:

Basically we have a digital DSD FIR filter implemented via discrete analog components.  It will to me always be a 1 bit FIR filter, until the day we get real multi-bit Delta Sigma source material to send to it :)

 

Yes, now we send it 1-bit data, but the converter itself doesn't care if we chose something else... :)

 

[Miska]

11 minutes ago, Andy-HandyHifi said:

You are correct that DSD does not bypass the segment DAC in the Burr Brown implementations in something like a direct mode.  BUT, it is really semantics because when DSD is present, those same segments are grouped together to form 8 taps. 

 

I meant that the TI/BB only has "direct mode", nothing else for DSD. While some others have switchable choice between DSP and non-DSP paths.

 

 

[Miska]

21 minutes ago, Andy-HandyHifi said:

I want to see what your filters I can come up with and how they compare to the filters on the iFi FPGA, ESPECIALLY when going from lower rate DSD to higher rate DSD.   Like DSD64 TO DSD1024.

 

At least you have multiple options for doing it. At the moment I'm personally sticking to DSD256 using ASDM7EC modulator. Since there are no computers available at the moment that could run this modulator at higher rates... (in real-time, of course with HQPlayer Pro you could run offline conversion to DSD1024 or DSD2048)