Choosing a DAC for DSD playback?

[CJH]

I want a sub $1000 DAC for DSD use. Will be upsampling files using HQP to 512. What DAC measurements are important to consider and compare for best DSD playback?
CJH

 

[Veri]

Both AKM and Sabre's newer chips should process native DSD streams as per HQP. Your main concern should be "pops" between switching PCM/DSD and other such DSD processing nasties. Besides that the general measurement suite should be indicative for both PCM and DSD.

 

[shanecoughlan]

Amir has great reviews. Check out the Topping NX4DSD and build from there. I got the NX4. Very happy.

 

[Veri]

Amir only measured one DAC with DSD-only architecture so far btw. It measured pretty decent https://www.audiosciencereview.com/...measurements-of-holo-audio-cyan-dsd-dac.6992/ (107dB SINAD when not converting from PCM, great multitone result, very low noise shaping)

But you don't need to spend so much since basically all chip-based DACs do DSD nowadays. For example the newly released/reviewed D70 would do DSD fine, but there are just about zero user reviews of that one so far. User reviews are pretty important for your (kinda niche) use with a dedicated HQPlayer set-up. You don't want a clicking/relay pop noise at the end of every song for example... you could also just order whatever you want with a good return policy (amazon, ...) and send it back if it doesn't work out of course!

Perhaps consider asking the same question on audiophilestyle forum. They will give you a lot of suggestions but they will all be very expensive lol. Hold on to dear wallet

 

[CJH]

Looking at the D70 and M300. Was curious what measurements Amir makes should be considered when looking for top DSD performance. I'm guessing 20 bit linearity doesn't matter but low noise/distortion does. Does the ESS hump matter for DSD? Am currently using an iFi Micro DSD BLK but curious if I can get better DSD sound for not too much $$.
CJH

 

[Ron Texas]

You better have a powerful computer as HQP uses lots of processing power. Also make sure you do not have room EQ needs.

 

[Calexico]

Why converting everything to dsd?
Do you think it will sound better ?

To my mind you should be sure because it's a complex solution to have a high power computer to convert in real time. Is you computer noiseless?

If you don't care using lot of cpu
Technically you can have more informations on 768khz 32 bit than on dsd.
So why not converting to 768khz 32bit?
New akm ess dacs with xmos can do it.

Not sure if the sound will be better than leaving the dac doing the oversampling.

 

[Ron Texas]

Why converting everything to dsd?
Do you think it will sound better ?

Not sure if the sound will be better than leaving the dac doing the oversampling.

This has been debated endlessly in other forums. I don't see any advantage to doing it. There is only so much information in Redbook. converting to high DSD rates doesn't create anymore. The argument is utilizing the processing power of a PC is better than using dedicated hardware in a DAC. I don't buy into that, but swing away to your delight.

 

[Calexico]

@CJH once i tried different oversampling plug ins and i found that the libsamplerate secret rabbit code was the best sounding to my hear. It's freely included in linux music softwares like amarock.
I don't know why nobody tries it.
I think it easier to convert pcm 44.1khz to 705.6 khz than converting to DSD.
That means that you can use the power of the cpu more efficiently.
It's better having a high quality upsampling than a middle quality of pcm to dsd convertion that takes more cpu.

 

[Ron Texas]

@CJH once i tried different oversampling plug ins and i found that the libsamplerate secret rabbit code was the best sounding to my hear. It's freely included in linux music softwares like amarock.
I don't know why nobody tries it.

Secret rabit takes a lot of CPU power, at least at the highest quality settings. I use it for non-realtime sample rate changes.

 

[CJH]

I started upsampling years ago using R8brain finding 24/96 sounded better than using the SRC in my Tact or DEQX. I use an older 4 core i7 with Windows 10 Pro & HQP. With 705/768 CPU is at 10%, with DSD 256 around 35% and can do DSD 512 with the smaller filters/modulators but just barely with CPU at 50+%. I prefer the DSD sound (sounds more natural) and am trying to find out what DAC measurements are important for best DSD performance.
CJH

 

[Eirikur]

You need to carefully look at the DAC chip if you want "true" DSD playback.
I've scanned several ESS datasheets and they all seem to transcode DSD into PCM first, here's an example of the block diagram for ES9038PRO:

Unless I interpret this diagram wrong, there are no separate paths for PCM and DSD. Note that this diagram is essentially the same for the ES9038Q2M which is used in the mentioned Topping NX4.

In contrast, you could look at the more detailed block diagram for the AK4497(EQ).
Here I do see explicit dual pathways to separate DSD from all PCM processing when selected:

My conclusion (and please correct me if I'm wrong!):

• ESS (any type) will always execute a (source)DSD->PCM->DSD->analog transformation chain
• AK4497(EQ) can do a direct (source)DSD->"DSD filter"->analog
  • AK4490 and AK4493 also have this but miss out on the "DSD filter"
  • AK4495(S) does not have a DSD path!

Links:

• ESS datasheets: ES9018 , ES9038PRO , ES9038Q2M (leaked document?)
• AKM datasheets

 

[Veri]

@Eirikur Interesting, however I've seen measurements where Sabre chips fed with upsampled DSD (using HQPlayer or other software) seem to have differing performance and measurements compared to PCM. Don't really remember where I saw that though, perhaps on CA forum. Maybe @Miska can chime in?

 

[Music1969]

ESS (any type) will always execute a (source)DSD->PCM->DSD->analog transformation chain

A popular belief is ESS convert DSD (1 bit SDM) to multi-bit SDM, for digital volume control... but the input sample rate itself is not changed (that's the belief).

Some people have shown the ESS chips measure objectively better when fed high DSD rates.

ESS chips have long been a bit of a 'black box'. It discussed here (see below link) a few years ago by some clever people and I'm not aware of any new information that has come out (regarding the ESS chips anyway).

https://audiophilestyle.com/forums/topic/19381-ess-sabre-and-dsd-volume-control/

Unless I interpret this diagram wrong, there are no separate paths for PCM and DSD.

This is part of the 'black box' that ESS have kept to themselves. They don't show absolutely everything that's happening in their block diagram.

Like you mentioned some AKM chips have a 'DSD Direct' mode. Not all DAC makers give this option to the end user though. RME do with ADI-2 DAC and Pro models.

 

[Music1969]

I want a sub $1000 DAC for DSD use. Will be upsampling files using HQP to 512. What DAC measurements are important to consider and compare for best DSD playback?
CJH

Jussi (HQPlayer dev) highly recommends the RME ADI-2, especially with 'DSD Direct' mode enabled via the front screen.

You'll need to do volume control separately though becuase 'DSD Direct' mode disables all volume control of the RME ADI-2.

 

[Eirikur]

A popular belief is ESS convert DSD (1 bit SDM) to multi-bit SDM, for digital volume control... but the input sample rate itself is not changed (that's the belief).

https://audiophilestyle.com/forums/topic/19381-ess-sabre-and-dsd-volume-control/

This is part of the 'black box' that ESS have kept to themselves. They don't show absolutely everything that's happening in their block diagram.

You're obviously correct, the diagram lacks too many details to be sure of anything. One thing I do know for sure from the (leaked) ES9038Q2M documentation is that ESS doesn't provide an interface for "DSD direct" or any thing like it. All other manufacturers I've looked at (TI/BurrBrown, CirrusLogic/Wolfson, AKM) do have explicit interfaces.

I followed up the link and also found an ESS whitepaper pointing to a specific US patent 7,058,464 that doesn't really make things clear to me, except for one (revealing?) picture:

They filter a multi-bit PCM input signal and upsample to sigma-delta to produce a PWM signal (also note feedback loop through 240). I could very well imagine them creating a similar PWM signal out of a DSD stream.

Sort of clarifying text from the patent:

In one example embodiment, the oversampling filter 232 may modulate the 16-bit filter output signal 225 at 44.1 kHz into a 4-bit oversampled signal 227 at 1.411 MHz (i.e., 32* 44.1 kHz, which is also called “32×” oversampling). In other embodiments, the oversampling filter 232 may modulate a wide-bit signal (e.g., 12–24 bits) into a signal of only a few bits (e.g., 2–6 bits).

Another figures show how this is integrated (no details about blocks 320 and 330):

The link you provided also gave some additional information as to the commonality of PCM and DSD processing (Andrew Allen Ballew):

1. The internal processing uses an ASRC which in turn incorporates a digital filter
2. The digital filter is claimed by ESS to work at 32 Bits
3. The digital Volume control is claimed by ESS to work at 32 Bit Precision
4. The digital filter and ASRC can be bypassed, if you do this DSD no longer works

Number 4 is very telling. The ESS does sample rate conversion on DSD. If you turn off the ASRC, bye bye DSD. This combined with their precision 32 bit volume control, and the digital FIR filter.
Strongly suggests DSD is converted to a common internal format, which is some form of PCM.

and also (indirect quote from Mark Mallinson of Resonessence Labs):

Inside the Sabre DAC chip the DSD stream is processed at the DSD rate, but it is extending to have a 32 bit representation so that it can be scaled. This is NOT the same as converting to PCM as that implies decimation filtering. Inside the Sabre, all DSD data is processed in a PCM domain, however its PCM at 64*44.1kHz = 2.8224MHz. This data is then applied to the modulator.

and finally Mytek's Michal Jurewicz

The signal path is DSD>32bit vol control in DSD domain>DSD LPF in DSD domain>modulator>6 bit DAC

In conclusion I have to nuance my opinion: it is very well possible that all signals are ultimately sigma-delta in the SABRE chips - perhaps judgement is also subject on how "pure" we consider the described PWM stage + error-loopback (which integrates the PWM again).
If the (anecdotal) additions are correct there is a indeed a "direct" DSD path without PCM decimation.

PS: My "vested interest" in researching this is my Topping D10 DAC which employs a single ES9018K2M.
I have no complaints about the quality but I'd like to understand what they are doing and whether or not I'm listening to a "true DSD" decoding chain.

 

[Music1969]

One thing I do know for sure from the (leaked) ES9038Q2M documentation is that ESS doesn't provide an interface for "DSD direct" or any thing like it.

Correct

If the (anecdotal) additions are correct there is a indeed a "direct" DSD path without PCM decimation.

Yes, 1 bit SDM is converted to multi-bit SDM (for digital vol control). But the input DSD sample rate is unchanged. Maybe...

I mentioned that thread is old. Maybe something has changed? Who knows. The ‘black box’ mystery continues

 

[bravomail]

You need to carefully look at the DAC chip if you want "true" DSD playback.
I've scanned several ESS datasheets and they all seem to transcode DSD into PCM first, here's an example of the block diagram for ES9038PRO:
View attachment 29960
Unless I interpret this diagram wrong, there are no separate paths for PCM and DSD. Note that this diagram is essentially the same for the ES9038Q2M which is used in the mentioned Topping NX4.

In contrast, you could look at the more detailed block diagram for the AK4497(EQ).
Here I do see explicit dual pathways to separate DSD from all PCM processing when selected:
View attachment 29961

My conclusion (and please correct me if I'm wrong!):

• ESS (any type) will always execute a (source)DSD->PCM->DSD->analog transformation chain
• AK4497(EQ) can do a direct (source)DSD->"DSD filter"->analog
  • AK4490 and AK4493 also have this but miss out on the "DSD filter"
  • AK4495(S) does not have a DSD path!

Links:

• ESS datasheets: ES9018 , ES9038PRO , ES9038Q2M (leaked document?)
• AKM datasheets

Can't help but laugh. All this talk about DSD, high bit rates DSD, and at the end PCM

 

[Eirikur]

Can't help but laugh. All this talk about DSD, high bit rates DSD, and at the end PCM

I expect no audible difference either way of decoding... but without the means to actually check this, how can we be sure?

In the mean time I've looked up the DACs of a large number of SACD players, and only McIntosh and Audiolab use ESS Sabre that I have seen.
McIntosh is rumored to prefer PCM over DSD (fair enough), and Audiolab disqualifies itself by implementing MQA in their 8300CDQ (i.e. no regard for audio integrity).

 

[diegooo1972]

Talking about DSD we are going to add mathematically just dynamic. Not sure what imply.
But you sure can't comapare PCM upsampling and dsd conversion.

Talking about PCM upsampling I don' t think any kind of upsampling can be usefull.
At best what I obtain must sound the same.
From what I understand of signals Nyquist is strictly. All the informations I need are in the double frequency of a signal.
Because over 22khz we can't hear and so the fourier transform over 22khz is null for the human ear ( 22khz is quite high for everyone at every age ).
He states that the minimal sampling frequency to avoid aliasing and loss of information is the double of the frequency I need to sample.
We can oversampling but we must consider that at 44.1khz we already have all the information we need without aliasing and loss of information about the 22khz range.
Every oversampling we do is going to interest only frequencies over 22khz and not correcting or changing frequencies below 22khz in any way.
Otherwise Nyquist were wrong. And of course he wasn't. You can prove that.
That seems strange but it's what happens.
At my age I can hear 15khz. In my case 30khz sampling will do. Let's say 32 or 34 at best.
In double blind test I can't distinguish between 44.1 96 and 192. I tought I can but that was not true.
What i'm going to do with upsampling is to change a little the wave and not going to create more informations about frequencies below 22khz.
I'm just going to add more informations over 22khz. But I just add more informations mathematically or artificially if you prefer.
Correct informations maybe but still Informations that I can't really listen cause they belong to a range of frequencies I can't really hear.
So if this considerations are correct what am I going to obtain ?
I must say that nothing is gonna change, whatever we think, by upsampling PCM for a human ear.