Adventures in Upsampling to DSD

[Ron Scubadiver]

First off the hardware I have available right now is an Intel NUC with an N3700 CPU and a Grace M9xx.  An N3700 has about 33% of the processing capability of a gen 7 i5 two core notebook CPU.  The Grace M9xx can do DSD 256 with the latest firmware upgrade.

 

With Jriver (on trial) the best I could do was DSD 64.  Going to DSD 128 caused skipping.  Next I tried foobar using the SACD input component and it's included DSD processor.  I can get DSD 256 without skipping.  CPU usage with SDM type B is 50%, but the system runs at normal temperatures without the fan ramping up.  That tiny fan can turn 8,000 rpm and if it gets there, you know it.

 

See this link for details on foo SACD input and SDM types:

 

https://diyaudioheaven.wordpress.com/digital/pc-software/foobar-2000-for-dummies-part-3-new-experimental-sacd-plugin-v-0-9-x/

 

SQ is better than no upsampling.  I am less certain if SQ is better than sending 352800/384000 PCM to the DAC upsampled with two instances of Sox Mod 2 set at their defaults.  CPU usage is minimal for PCM output. 

 

Here is the part that I have trouble understanding.  Upsampling to high PCM rates is supposed to bypass upsampling and filtering by the DAC.  The argument goes a computer can do this better with software because it has more resources.  Upsampling to DSD 64 bypasses even more of the DAC which normally turns PCM into something like DSD.  DSD 128 should provide the  additional benefit of moving ultrasonic noise further away from the audio band.  I don't have any idea why DSD 256 would be better or if I had compatible hardware DSD 512.

 

Someone around here noted that some recording studio said they were getting better results with a DSD 255 recording machine, even on child copies at 24/96 and redbook.  That's well and good for recording and down sampling because the process starts with more data.  However, upsampling redbook creates no new data.  It just sends more interpolated data to the DAC.  It would seem all that would accomplish is making all the hardware work harder.

 

There are other issues.  the DSD conversion algorithms are from Phillips.  They are obviously at least twice as efficient as whatever Jriver is using.  I suspect these routines use much less processing horsepower than HQPlayer judging from the relatively powerful computer hardware that group of audiophiles is using to reach DSD 512.  An i5-7600 has around 6 times the processing power of my N3700.  Perhaps HQPlayer and the DSD processing hardware of high end DAC's may have some secret sauce.

[Ron Scubadiver]

15 hours ago, Dr Tone said:

 

Yes this is true.  With the excellent upsampling/filters/modulators in HQPlayer you can get great results.  With JRiver not so much.  I wouldn't even bother turning it on is JRiver.

OK, assuming one is using HQPlaayer, why would DSD 256 or DSD 512 sound better than DSD 128?  Is it just more is better?

[Ron Scubadiver]

I have found an objective reason to use DSD 256 instead of DSD 128 on my foobar based setup.  With DSD 128 there is a double click when starting playback.  That changes to a single click of lower amplitude with DSD 256.  My understanding is this behavior is common, although some combinations of DAC's and players probably can avoid it.

[Ron Scubadiver]

Perhaps I lack the desire, or I can't hear that well.  I am on the old side.  With the hardware I have, upsampling in software was no better than choosing a different filter which is possible on the Grace M9xx.  I tried HQPlayer a bit.  It doesn't suit me.  Perhaps there is a way to use it with a difference interface and with some DAC's it really sings, but I don't have time or money for that.

 

[Ron Scubadiver]

After a great deal of experimentation I found Foobar using Sox Mod2 with HQ, minimum phase and 90% bandpass produced the best results.  I couldn't get Jriver to sound right no matter what.  It may have been that I am using a DSP to boost at 40hz.  I must be one of those people who does not like pre-ringing.  Note that the Grace M9xx has 4 filter options and minimum phase slow roll off is one of them.  In the documentation for the AK chip there are specs for the filters going up to 192k, so I believe it is reasonable to believe at 352.8/384 there are no hardware filters.

 

One thing I have noticed is JRiver has a sort of antipathy towards upsampling with their fixed parameters on Sox and repeated statements that they will not implement options.  That seems kind of strange for a playback program with so many options.

 

Foobar's main drawback is it is kind of a kit.  Lots of assembly is required.

[Ron Scubadiver]

15 hours ago, buonassi said:

 

To me the benefits of upsampling have been diminished with my new DAC - which also has selectable filters.  So it's not just you.  I've read about the M9XX and almost bought it for this very reason but I went with the iDAC2 by iFi.  As OP stated, if your software can do a better job of filtering and interpolation than the DAC, then it becomes easier to hear the increase in SQ when you bypass its filters.  The main benefit for me, is that I don't care for the lively ringing, harmonic excitement etc of traditional filters, and I'm able to dial in my filter parameters to my liking, controlling total overall ringing with the slope of the rolloff, as well as where that ringing takes place (pre or post transient) with a phase slider.  

 

 

It's definitely YMMV.  BTW, the Grace suffers from pops and clicks when playing back DSD, and especially if switching back to PCM.  This is one of the things they don't tell you in the reviews.  

[Ron Scubadiver]

4 hours ago, buonassi said:

Wonder if it's a usb cable issue? My idac2 so far has been problem free.  There's the usual clicks when you first start playback or skip a track, but after that, its all smooth and no pops even when the tracks change by themselves. 

I have noticed with foobar upsampling to DSD 256 produces a barely audible click.  If I set the player up so that everything goes to DSD, there are no clicks when reverting to PCM, because it doesn't happen.  I am not sure if SQ is improved.

[Ron Scubadiver]

4 hours ago, Miska said:

 

You need to compare different conversion algorithms, because with a delta-sigma DAC the algorithms define large portion of the quality/performance...

 

Well, Foobar uses something from Phillips called SDM and I used SDM B out of the 4 choices, A, B, C and D.  Chances are it isn't as good as what you are using and my computer is resource constrained having only an N3700 processor.  I am getting the best results with the minimum phase, sharp rolloff setting available on the M9xx and letting the DAC do all the work.  It isn't all that sharp of a rolloff, about 90/110.  This is what Bryston BDA-3 uses and it has the AK chip too.

[Ron Scubadiver]

25 minutes ago, Miska said:

 

With AKM chips it is also important to pay attention to check whether the chip is in "DSD Direct" mode. For example RME ADI-2 and TEAC UD-503/NT-503 have this feature.

 

I don't know if my Grace M9xx has that feature.  If it does, it would have to be by default because there is no option for it on the setup interface.

[Ron Scubadiver]

38 minutes ago, Miska said:

 

You could ask Grace how they've programmed DSD mode to work. But since according to their documentation the volume control is digital, then it very likely isn't in DSD Direct mode because that would disable the digital volume control... So very likely the DAC isn't really suited for DSD use but instead could benefit from upsampling in PCM domain and being fed at highest possible PCM rate.

 

I think your key statement is "DAC isn't really suited for DSD use..."  I may give upsampling to 384k another shot.  The N3700 has more than enough capability to handle it with SOX.  It will have to wait for tomorrow as I am out for the weekend.