A novel way to massively improve the SQ of computer audio streaming

[hols]

21 hours ago, Gavin1977 said:

Great feedback @hols  what do you mean by the above sentence... do you mean DSD still sounds better than PCM upsampled at the same rate?  Is the difference large enough to warrant the huge processing requirements for DSD?

 

I ask because I run T+A DAC 8 DSD, I can’t do NOS PCM, and am presently CPU limited (like most) to DSD256 for demanding filters.

 

Actually it's difficult to describe the PCM1411.2. I can try to explain in the following manner. I tend to think that from 44.1 to 705.6 the improvement is along the same line, i.e increase in detail but without much change in character. But reaching 1411.2 there seems to be a change of character of the SQ. The most obvious is you tend to focus on individual instruments because they are so real and full of feelings so much so that you won't be paying too much attention to the whole orchestra. It is like you see more of the tree but not the forest as a whole. The details of the tree is portrayed beautifully and as musical as DSD. The music phrases comes out beautifully, the change in tone also brought out beautifully and there is not much air between the different instruments. In DSD I would think the image is larger and feels more juicy, a bit more harmonics but the flow as not as swift as in PCM.

To me this is a new experience and I think it may need maybe some specific filters to further bring out its benefits.

 

If I can start again from the very beginning I would probably go for PCM upsampling rather than DSD because the demand on computer is much less and there is a high potential to get equal satisfaction on musical feeling as in DSD. YMMV.

[ASRMichael]

So we are back to two box solution discussions. The Extreme is effectively a two both solution. So 2 high powered servers (same spec). Would have been a fairer test against the Extreme. 
 

I suspect 1 year ago when diy members said 1 box is the best solution was because they did not have two similar spec boxes. 

[ted_b]

7 minutes ago, hols said:

If I can start again from the very beginning I would probably go for PCM upsampling rather than DSD because the demand on computer is much less and there is a high potential to get equal satisfaction on musical feeling as in DSD. YMMV.

Hols, sorry to be so dense here, but are you saying that, with the endgame being PCM 1411.2k, any source files, even DSD files?  IOW, you are saying DSD64 to PCM 1411.2k is better than DSD64 to DSD256 (using EC modulators), not to mention the cpu savings?

[hols]

17 hours ago, sig8 said:

I am considering T+ and Gaia, but based on @hols reporting above if higher speeds in PCM and perhaps in DSD sound better, then Gaia will limit you to DSD512, and PCM to 768 kHz. Don't know if T+ at higher speed is better than T+ and Gaia combo at lower speed. Thoughts? Experience?

I don't have personal experience on Gaia. All along my focus is on USB rather than I2S. I have tried Audiobyte, and Singxer SU-1 with Holo Spring but I don't prefer using them but I must say that I have not modified the SU-1 or tried good enough cables. I have not tried I2S input for Terminator Plus. But Terminator plus performance with USB is already very satisfying be it in PCM 705.6 or DSD 256, not to say if you go all the way to PCM1536. I would suggest you go for Terminator Plus first. You can always add a Gaia afterwards. 

I may be able to get hold of a Gaia sometime later. I will report back if I have further personal observation.

[matthias]

@hols

 

Just for clarification:

When you are talking HQplayer+NAA with the Taiko Extreme which version of HQplayer do you use? Embedded?

Thanks

 

Matt

[hols]

8 minutes ago, ted_b said:

Hols, sorry to be so dense here, but are you saying that, with the endgame being PCM 1411.2k, any source files, even DSD files?  IOW, you are saying DSD64 to PCM 1411.2k is better than DSD64 to DSD256 (using EC modulators), not to mention the cpu savings?

This is too difficult for me because I have not tried converting DSD64 to PCM 1411.2. I can only say is for 44.1 files upsampling I would probably prefer the sound of PCM1411.2 than DSD256 EC modulators. And if that's the case then most likely DSD64 files may come to the same result. 

But I must reiterate that PCM 1411.2 is a new experience and I cannot confidently say that everyone will like it. You really have to listen to it first before making decision.

 

[hols]

2 minutes ago, matthias said:

@hols

 

Just for clarification:

When you are talking HQplayer+NAA with the Taiko Extreme which version of HQplayer do you use? Embedded?

Thanks

 

Matt

HQplayer 4 desktop

[ted_b]

Hols, thanks. I have always preferred PCM files staying PCM (and DSD staying DSD) with my Holo dac, for example. So your conclusion that even higher resolution 1411.2 sounds great with red book fits my expectations. I’d be very surprised, however, to hear that DSD to 1411.2 PCM ( an inherently lossy conversion) upsampling withstands and is superior to DSD remodulation. Looking forward to more experiments. 😀

[matthias]

12 minutes ago, hols said:

HQplayer 4 desktop

 

Would HQPlayer Embedded be possible with the Extreme?

Thanks

 

Matt

[ray-dude]

Alas, no when running Windows.  HQPE requires linux.   You can boot the Extreme to HQPOS on a USB stick (or NAA) and use the hardware without all the OS tweaks.  I found this did not match the SQ of running HQPDesktop in the tuned Windows config on Extreme (not even close...there is a LOT of secret sauce in the OS tuning on Extreme)

 

 

 

 

[Exocer]

3 hours ago, ray-dude said:

Alas, no when running Windows.  HQPE requires linux.   You can boot the Extreme to HQPOS on a USB stick (or NAA) and use the hardware without all the OS tweaks.  I found this did not match the SQ of running HQPDesktop in the tuned Windows config on Extreme (not even close...there is a LOT of secret sauce in the OS tuning on Extreme)

 

Thanks @ray-dude.

 

@guiltyboxswapper - I am grateful for your nudge to research HQPlayer. It makes for an interesting combination with the Yggy and I look forward to playing with the filters more. The pair sound very good together! There were some hurdles with HQPlayer Embedded initially since some of the default settings did not align with the documentation (on GentooPlayer) but once those issues were resolved I was able to locate the configuration page.

 

Other recent changes:

 

1. Switch from HDPlex 400W DC-ATX to 800W DC-ATX.

• I think this made for a rather worthwhile improvement even with my current SMPS powering the server.

2. External clock for ER. Thanks to @MartinT's experiment I gave this clock a try and I do feel it improved the soundstage, separation and calmness without a loss of dynamics. This is no Mutec Ref 10 but for a fairly cheap price this experiment has proven to be rewarding. I have since switched my LPS 1.2 to the clock (since I prefer stock SMPS over the LPS 1.2 for the ER).

• 
• 

 

Happy listening all,

-Rob

 

[sig8]

Could not resist any longer, T+ and Gaia on order. Will compare with Holo Spring Kitsune edition + Singxer SU-1 modified) combo. Thanks everyone for their input.

[cat6man]

18 hours ago, Exocer said:

 

Thanks @ray-dude.

 

@guiltyboxswapper - I am grateful for your nudge to research HQPlayer. It makes for an interesting combination with the Yggy and I look forward to playing with the filters more. The pair sound very good together! There were some hurdles with HQPlayer Embedded initially since some of the default settings did not align with the documentation (on GentooPlayer) but once those issues were resolved I was able to locate the configuration page.

 

Other recent changes:

 

1. Switch from HDPlex 400W DC-ATX to 800W DC-ATX.

• I think this made for a rather worthwhile improvement even with my current SMPS powering the server.

2. External clock for ER. Thanks to @MartinT's experiment I gave this clock a try and I do feel it improved the soundstage, separation and calmness without a loss of dynamics. This is no Mutec Ref 10 but for a fairly cheap price this experiment has proven to be rewarding. I have since switched my LPS 1.2 to the clock (since I prefer stock SMPS over the LPS 1.2 for the ER).

• 
• 

 

Happy listening all,

-Rob

 

 

@rob

@ray-dude

@hols

@austinpop

 

have you experimented with the hqp buffer time?

i have found that increasing it always improves the SQ.

 

the maximum permitted through the hqp web interface is 250ms (range checking on input**)

with an opticalRendu, i found better SQ with 500ms, which could be set by editing the settings.xml file

also, with a NUC (audio-linux for example, haven't tried w/ euphony) which has a lot more RAM available,

you can increase the buffer time even much more significantly (just note that the buffer is in units of time, not bytes,

so that a larger value is possible with red book than with hi-rez sources)

 

i'd be interested in any conclusions you have with both the extreme and diy servers w.r.t. buffer time settings.

 

**interesting story--------i had a vp who was getting a demo of a new modeling program for laying out the location and design

parameters for mobile cell sites, incorporating terrain, propagation, etc. 

he sat down and immediately entered a value for a cell tower that was 2000 miles tall, the program crashed, and he walked away saying their program design had flaws.  in my case, i set the buffer value to use the maximum amount of ram available, calculated using red book cd parameters.  when i went to play a 24/96 track, i crashed the NUC.  however, the maximum ram buffer did improve SQ.

 

 

[ray-dude]

When I was running HQP on my NUC, my point of diminishing returns on buffer was between 2000-3000 msec (set in the settings.xml).  Any larger than that and I wasn't hearing a benefit.

 

With Extreme, things are all different.  Buffer size has a definite impact on dynamics.  Smaller the buffer, the better the dynamics but perhaps thinner?  Larger buffer slower but perhaps fuller.  The combination of DAC bits and buffer are basically tone controls now (on Extreme), and the final tweaking I do after adjusting some other part of the chain.

[StreamFidelity]

I think saving in RAM and Buffer size are two different things. Saving in RAM is about separating the music files from the network noise. The music is played from the fast working memory.

Buffer is about fetching the largest possible data packets on a slow network to avoid dropouts. In my experience, however, a large buffer has a very negative effect on the latencies. I have set both the network card adapter and USB to the smallest buffer.

See Denafrips USB ASIO as an example. The output latency is very low 1.50 ms. That makes an audible positive difference in my system.

 

[dminches]

36 minutes ago, StreamFidelity said:

I think saving in RAM and Buffer size are two different things. Saving in RAM is about separating the music files from the network noise. The music is played from the fast working memory.
 

 

It isn’t exactly the same thing but I run Euphony and Ram root.  All the my music is buffered and played from memory.  I have tried removing the network cable after the music is buffered and I hear no difference in sound quality.  This approach in Euphony does a great job in reducing network noise.

 

[cat6man]

36 minutes ago, StreamFidelity said:

I think saving in RAM and Buffer size are two different things. Saving in RAM is about separating the music files from the network noise. The music is played from the fast working memory.

Buffer is about fetching the largest possible data packets on a slow network to avoid dropouts. In my experience, however, a large buffer has a very negative effect on the latencies. I have set both the network card adapter and USB to the smallest buffer.

See Denafrips USB ASIO as an example. The output latency is very low 1.50 ms. That makes an audible positive difference in my system.

 

when running a one box solution and running in ramroot, the amount of ram limits the buffer size

[StreamFidelity]

24 minutes ago, dminches said:

This approach in Euphony does a great job in reducing network noise.

 

I like to believe that.

 

Still, you could take a look at your network map. There are a multitude of settings in the network adapter that influence the buffer in the data traffic. That's what I mean. There a small buffer ensures low latencies and, in my and other experiences, this improves the audio playback considerably. Look for example to Pink Faun: "During the process of developing the streamer 2.16, we found out that latency is one of the key elements to achieve the really smooth and natural sound; the lower the latency, the better."

 

My contribution should only make it clear that not all buffers are the same. I also think it makes sense to cache it in RAM, and the buffer should be very large there. The buffer in the network traffic, however, should be as small as possible. 😉

[dminches]

9 minutes ago, StreamFidelity said:

 

I like to believe that.

 

Still, you could take a look at your network map. There are a multitude of settings in the network adapter that influence the buffer in the data traffic. That's what I mean. There a small buffer ensures low latencies and, in my and other experiences, this improves the audio playback considerably. Look for example to Pink Faun: "During the process of developing the streamer 2.16, we found out that latency is one of the key elements to achieve the really smooth and natural sound; the lower the latency, the better."

 

My contribution should only make it clear that not all buffers are the same. I also think it makes sense to cache it in RAM, and the buffer should be very large there. The buffer in the network traffic, however, should be as small as possible. 😉

 

Understood, but if I don’t hear any sound improvement with my server off the network then what improvements could I hear in changing my network?

 

I also don’t know how to change the network adapter setting with Euphony.  I have done that when I run Roon in Windows.

 

[hifi25nl]

Audio latency (like in Audio pro) is for example the delay between hitting a key on a midi piano and the sound output.

This is of the order of milliseconds. 

If you want a very low audio latency, you need a powerful computer to obtain values even less than 1 ms.

This is important in live playing, you cannot hit a key and wait... Some have noted that the sound is a little different if audio latency is lower, but maybe this is because of the more heavy load on processors.

 

Is this important for Audio reproduction? Probably not. What I think is important is the variation of CPU latency, that is the time between a command and its execution. This is of the order of microseconds. This is where realtime kernels play an important role.

With the right processors you can achieve an average value of 3-5 microseconds, only dreamed of by a computer with Windows, however optimized.

 

So I would use high buffers (in a general sense), since the "audio delay" is not important but a high value buffer give the best framework to obtain a very low variation on CPU latency.

 

[onlychild]

21 hours ago, cat6man said:

 

@rob

@ray-dude

@hols

@austinpop

 

have you experimented with the hqp buffer time?

i have found that increasing it always improves the SQ.

 

the maximum permitted through the hqp web interface is 250ms (range checking on input**)

with an opticalRendu, i found better SQ with 500ms, which could be set by editing the settings.xml file

also, with a NUC (audio-linux for example, haven't tried w/ euphony) which has a lot more RAM available,

you can increase the buffer time even much more significantly (just note that the buffer is in units of time, not bytes,

so that a larger value is possible with red book than with hi-rez sources)

 

i'd be interested in any conclusions you have with both the extreme and diy servers w.r.t. buffer time settings.

 

**interesting story--------i had a vp who was getting a demo of a new modeling program for laying out the location and design

parameters for mobile cell sites, incorporating terrain, propagation, etc. 

he sat down and immediately entered a value for a cell tower that was 2000 miles tall, the program crashed, and he walked away saying their program design had flaws.  in my case, i set the buffer value to use the maximum amount of ram available, calculated using red book cd parameters.  when i went to play a 24/96 track, i crashed the NUC.  however, the maximum ram buffer did improve SQ.

 

 

I've spent several months messing with these settings on the Extreme and agree with Ray Dude that they are basically tone controls when using 10ms - 250ms,  until I tried "Default" on both settings (buffer time and bits).  To me, "Default" ended up being the best parts of all the other settings with none of the weaknesses.  

 

When using any setting between 10-250ms, I would prefer one over the other depending on the album I was listening to or because of some recent system upgrade or change.  "Default" has sounded good across all my music and through all the recent changes and upgrades.  It was a very obvious improvement, not something I had to focus on to hear.  

 

When ever I would increase the buffer, let's say from 20ms to 50ms, the top end sounded rolled off, there was a loss in dynamics, loss in details and speed, but sound was fuller and smoother.  With Default - I am hearing all the details, speed, fullness, dynamics, etc with none of the drawbacks.  

 

Wanted to see if anyone else is hearing the same thing with "default" that I am so I can finally put this dead horse to rest.  

 

I'm running Extreme to Dave to AFC-10 to Susvara.  

[lmitche]

Is it just me or do any of you also notice a SQ improvement with the latest version of Roon (build) 610?

 

It looks as though this version is distributing processing across more cores which could explain what I am hearing.

 

Larry

[lmitche]

As promised, below please find an update to my project list posted on July 4th.

------------
Project goal: There are a lot of great online posts about which SFPs sound best. Which sound best here?

Prior Results: Here is my ranking from best sounding to worse:

 

1) Finisar 10GB/s LR LW 1310nm SFP+ Transceivers Single Mode
2) Finisar 10Gb/s LR LW 1310nm SFP+ Transceivers Single Mode (Wide Temp)
3) Fs.com - Intel E10GSFPLR Compatible 1000BASE-LX and 10GBASE-LR SFP+ 1310nm Transceivers Single Mode

 

Latest results: No major change in the above, but I did observe that mixing wide temp and regular Finisar SFPs sounds better than wide temp to wide temp or some wide temp to regular temp Finisars. I am beginning to think there are variations between individual modules that are more important than the temperature rating.

Next steps: Done for now
------------
Project goal: Find a replacement for the unobtainable Apacer DDR4 2666mhz UDIMMs that matches or exceeds Apacer SQ. Best case would be another Samsung B-die memory module or whatever variant of Samsung memory chips that replaced B-die.

Latest results:

There are a plethora of Samsung B-die memory modules available from the major memory packaging and distribution companies. Take a look here for the "google" of Samsung B-die.https://benzhaomin.github.io/bdiefinder/

 

I have tested 3200, 3600, and 4133 mhz modules from Team Group and Gskill in the past few weeks. Despite all being B-die, there are differences in SQ between these modules. The more expensive "higher" binned parts do indeed sound best even when running below their maximum rated speeds. All of the above beat the Apacer 2666mhz SQ.

Next steps: Done for now
----------
Project goal: Can powering and controlling an active cooling solution from a separate power supply instead of the motherboard minimize SQ impact?

Prior results: Noctua speed control with a 140mm Noctua fan and heatsink seems to have no impact on SQ. Fan speed can be adjusted for mechanical silence. Any mechanical vibration does not appear to impact SQ. Active cooling with a fan does lower operating temperature from 60 to 42C.

Latest results: SQ is not impacted by the use of an externally powered fan.  Caveat - fan power-on/off and speed control is manually controlled with this solution and can be an inconvenience and/or risk to the hardware if not powered up with the cpu and motherboard.

Next steps:

Test active AIO pump and fan combos with high power Intel based Hqplayer solution. Need to find a fan and pump combo where it is possible to control the pump speed as well as fan speed from one speed control. Are one, two or three fans best? Is pump vibration a factor in SQ?
------------
Project goal: Can the Monoprice fiber USB 3.0 extension device endpoint be powered directly creating a galvanically isolated solution?

Prior results: A dead cable was dissected. Circuit board markings indicate that the copper cable carries 12 volt boosted power from the host end of the cable.

Latest results: Work has stopped on the Monoprice project as a different fiber solution for galvanic isolation of the server from the DAC and network is available. An Adnaco fiber USB rig with the single mode Finisar SFPs described above beats the SQ of the Monoprice solution.  Here the Adnaco remote USB endpoints connect the DAC and network gear, and high quality 5 volt LPSes, (LPS1s and PH SR-4s) power the Adnaco endpoints. The combination of the Adnaco technology, single mode fiber and clean DC power yields terrific clarity and musicality.

Next steps: Done for now
------------
Project goal: Explore possibility of changing the host environment variables to accommodate recent changes to Roon and thereby improving SQ. Recently there has been much discussion of this in the Taiko Audio Extreme world.

Questions: A comparison of various Roon versions running on ram rooted AL shows a growing memory footprint over time. Current servers are configured with 8gb of RAM. Does 16gb of ram lead to improved SQ? Are other improvements possible?

Prior results: Total VM footprint of Roon 1.71 measures close to 8gb with a 2.5 gb working set. With a ram based root drive this may be causing unnecessary IO.

Latest results: No difference in SQ was observed between 8gb(2x4gb) of memory and 16gb(2x8gb) of memory. Nevertheless switching to 16gb (2x8gb) as a new standard makes sense as 1) 4gb modules are getting more difficult to find, and 2) 16gb is a more future proof solution by leaving room for expansion.

Next steps: Done for now
------------
Project goal: Test SQ with a high quality SMPS ATX. With all the recent improvements is an LPS still worth the bother?

Latest results: Two power supplies were tested, the first an EVGA 1600 ATX power supply and the second a DELL 300 watt laptop SMPS in combination with a HDPLEX 400 watt DC to DC ATX with EPU/CPU power to the motherboard.  Both configurations created an enjoyable presentation but neither was as good as single or dual LPSes at 12 and 19 volts. The Dell Supply edged out the EVGA with very listenable presentation. 

 

If one didn't know better, either supply delivers a musical and  enjoyable presentation. But yes, a dual LPS solution sounds best and is a worthwhile upgrade.

Next steps: Done for now
------------
Project goal: Test SQ of the new 300 watt Hdplex LPS

Latest results: In July, I managed to secure a new 300 watt Hdplex LPS for a new customer build.

Background: I have two of the prior 200 watt models here. Using the two 200 watts, one for a 19 volt feed to a Hdplex DC to DC ATX and the second for the 12 volt cpu power is an improvement over a single 200 watt feeding both voltages. Splitting the power across two supplies results in a terrific SQ bump.

 

Good news, the new Hdplex 300 watt feeding both 12 and 19 volts power to my server is very close in sound quality to the dual 200 watt configuration. I am not sure I could tell the two configurations apart in a blind test. Clearly the increased power capacity and design innovations in the new supply deliver.

Next steps: The new model sold out quickly. When availability normalizes I look forward to testing two of the 300 watt models to compare with a single LPS.

 

I really like the Hdplex LPS solutions for their sound quality, versatility and value for money.
------------

Let me know if you have any questions.

 

Thanks,

Larry

[ray-dude]

3 hours ago, lmitche said:

Is it just me or do any of you also notice a SQ improvement with the latest version of Roon (build) 610?

 

It looks as though this version is distributing processing across more cores which could explain what I am hearing.

 

Larry

 

Emile noted the same over at WBF.  Maybe after all the bluster they're paying some attention to SQ? (hope so)  I've been holding off upgrading out of being in a happy music place right now (tweaking hiatus, at least for things not easily reversible)

 

Thank you for the project update Larry!  I learn a lot every time you post one of these, very much appreciated.

[guiltyboxswapper]

1 hour ago, lmitche said:

An Adnaco fiber USB rig with the single mode Finisar SFPs described above beats the SQ of the Monoprice solution. 

Do you mind sharing which model Adnaco?  There's quite a few.  As mentioned the JCAT XE + Monoprice aren't a stable option.  

 

And thanks for the memory tip, gskillz 4266mhz b-die running at 3200mhz (and standard! spd profile otherwise) is a good step forward indeed.