Best CPU for hqplayer

[chipvn]

On 1/21/2021 at 6:22 AM, Zauurx said:

I bought from a member of this forum, an HDPlex H3 with an i5-8400 on a Gigabyte Z370N motherboard.
After a quick optimization of Win10 LTSC with A0, PL .. MC 100, without using Fidelizer, I read the Redbook in DSD256 ASDM7EC without problem.
I can't seem to get the latency down. I don't know if the processor has a problem. No problem with the audio, only LatencyMon gives me bad results...
Or if I haven't found the correct settings in the bios.
Even if the optimization is a little brutal (I have not yet put my hands in the registry), this is what it gives with a set that is almost 2 years old.

52° after one hour of music 😉

 

Hi,

 

Are you able to do 24/192 to DSD256 using ext2 and ASDM7EC on this passive cooling system?

 

Thanks

[Zauurx]

I continued to optimize and find the solution to lower the latency (HPET off).
With EXT2, it's limit for 24/192 .. dropout every 5 sec.
No problem with a "2s" filter (eg short-lp-2s).

 

With EXT2 :

 

 

With sync-short-lp2s

 

 

 

[Miska]

44 minutes ago, Zauurx said:

I continued to optimize and find the solution to lower the latency (HPET off).
With EXT2, it's limit for 24/192 .. dropout every 5 sec.
No problem with a "2s" filter (eg short-lp-2s).

 

 

 

You could save some load especially at higher source rates by disabling Roon DSP and letting it pass data bit-perfect to HQPlayer.

 

Since you are playing local FLAC and and HQPlayer already supports volume leveling metadata in files (assuming that is what Roon is using), you could have same functionality in standalone HQPlayer and save more CPU time by dropping Roon out altogether.

 

[chipvn]

@Zauurx: Thank you for the info.

 

Planning to switch from ROON/ROON bridge configuration to ROON/HQPlayer/NAA configuration.

 

I'm still could not decide to either get a Mac mini M1 (16GB RAM) or build a 65W TDP passive cooling computer using something like HDPlex H3 V3 case. The reason why I want to go passive is not just because of fans noise but also dust issue which is quite common on active cooling system.

 

That's the reason why I'm trying to read around to see if there is any 65W TDP (none K version CPU) Intel solution that work for 24/192 to DSD256 (x44.1) with ext2 & ASDM7EC.

[Zauurx]

I used the same Roon> HQP on my oldie Lenovo Tiny M92 (i5-3470T)  (with 2 ethernet cards on the same machine) without difficulties to DSD128.
I bought this machine already mounted and the choice of I5-8400 seems good (no hyper-threading support. 6 cores)

You have to ask Miska and other users, apart from already having files in 24/192,
I'm not sure there is a difference in terms of SQ between 44.1> DSD256 and 192> DSD256.

As I already wrote, I tested a Lenovo I7 .. but could not silence the fans.
I took time to find the right optimization solutions for win10, which I would not do for IOS.

 

And I forgot, I have a convolution activated through Matrix menu in HQP. Without it is even better for the CPU load.

@Miska the only DSP under Roon is actually the volume adjustment between songs and albums (adjusted to -18lufs).
I don't know if HQP uses the same data. (R128 not Replaygain).
https://help.roonlabs.com/portal/en/kb/articles/volume-leveling
And Roon's total CPU load is ridiculous (screenshot same file 24/192 EXT2).

I will still test 😉

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

[Miska]

1 hour ago, Zauurx said:

And I forgot, I have a convolution activated through Matrix menu in HQP. Without it is even better for the CPU load.

@Miska the only DSP under Roon is actually the volume adjustment between songs and albums (adjusted to -18lufs).
I don't know if HQP uses the same data. (R128 not Replaygain).
https://help.roonlabs.com/portal/en/kb/articles/volume-leveling

 

It increases amount of data transferred between Roon and HQPlayer quite a lot, to avoid precision loss.

 

You can have ReplayGain metadata set based on any kind of calculation you like. I've been using loudgain myself, see also repo link here for some more details. I think I packaged it for Ubuntu as well. So it is based on R128 as well.

 

[Yviena]

Just curious but what is the per core usage for you guys running 10850-11xxx series with EC7 DSD256.

[Miska]

3 hours ago, Yviena said:

Just curious but what is the per core usage for you guys running 10850-11xxx series with EC7 DSD256.

 

[Miska]

Just tested also on 10850K + RTX3080 and for example sinc-L offloaded + ASDM7EC works fine.

 

[Yviena]

4 hours ago, Miska said:

 

Interesting  is that core 0 + core 5 SMT?

 

The cpu utilization  per core looks close to what I get when manually assigning 2 cores only to HQplayer coupled with multi DSP ticked on around 70-80% per thread.

[Zauurx]

16 hours ago, Miska said:

 

It increases amount of data transferred between Roon and HQPlayer quite a lot, to avoid precision loss.

 

You can have ReplayGain metadata set based on any kind of calculation you like. I've been using loudgain myself, see also repo link here for some more details. I think I packaged it for Ubuntu as well. So it is based on R128 as well.

 

If I want to understand. With ReplayGain or Loudgain, you have to retag your entire library and it doesn't work with Tidal (or Qobuz).
It's in Roon's interest to use his database and retag on the fly (but without modifying the files... alas).

[Miska]

8 hours ago, Yviena said:

Interesting  is that core 0 + core 5 SMT?

 

It is supposed to be 0 + 5. Windows is a bit of nightmare always. It doesn't always honor thread affinities exactly.

 

So if you want nice behavior, just stick to Linux.

 

8 hours ago, Yviena said:

The cpu utilization  per core looks close to what I get when manually assigning 2 cores only to HQplayer coupled with multi DSP ticked on around 70-80% per thread.

 

You practically see just the modulator load. ext2 filter is running on 6 cores, but it is so light that it is not much visible.

 

[Miska]

2 hours ago, Zauurx said:

If I want to understand. With ReplayGain or Loudgain, you have to retag your entire library and it doesn't work with Tidal (or Qobuz).

 

loudgain writes ReplayGain 2.0 metadata. There's no "ReplayGain" software, it is only specification of tag names where the dB adjustments are stored. It is totally up to the tagging software to decide how it comes up with the dB values.

 

If Tidal/Qobuz don't provide it in metadata, player cannot do leveling, especially not album leveling because it is streaming service, so you don't have visibility of larger entities.

 

2 hours ago, Zauurx said:

It's in Roon's interest to use his database and retag on the fly (but without modifying the files... alas).

 

Maybe I'll add it to HQPlayer's library "database" too. But this is one of the things you cannot do "on the fly", because for album gain (that makes most sense so that you don't destroy album's inter-track levels), or track gain, you need to read though the entire album/track.

 

One of the things I don't like in Roon is that it does non-playback related background activies on library and such during the playback. This is why I always want Roon in a different computer than HQPlayer. HQPlayer only goes through library files when you are scanning the library. Never during playback or otherwise randomly in background.

 

[wolft]

Hello, which of the new Z590 motherboards from Asus or Gigabyte are recommended for a new music server.
Which RAM should be taken?
The CPU should be an I9-11900k. Does this CPU make sense. Until then, an I9-10850K will be used.
Thanks in advance
Wolft

[luisma]

On 2/4/2021 at 5:49 AM, Miska said:

One of the things I don't like in Roon is that it does non-playback related background activies on library and such during the playback. This is why I always want Roon in a different computer than HQPlayer.

Last night I followed your advice @Miska

I moved my Roon server out of my HQPe server. Recently build a home NAS with TrueNAS virtualized under VMWare ESXi on a Ryzen 3700X, this ESXi is running TrueNAS with 24TB of available space, Windows Server 2019, 3 Centos Virtual Servers running Ansible and stuff and 1 Roon Ubuntu 20.04. I installed Roon and started streaming to my HQPe server (on which I disabled Roon). I wasn't expecting any MAJOR improvements (in terms of resources not SQ) but contrary to what I thought HQPe does improve from been isolated as a single server. The VM running Roon streaming 192khz content from Qobuz max CPU usage with only 2 cores assigned (which is recommended for the way ESXi handles CPU load distribution, if you need more power assign 3 cores, not 4 cores as 4 cores will throttle down your speed, so 3 is better than 4) never went over 22 percent. 256 DSD ASDM7EC streamed without a single hitch (except for having to manually restart every 30 min due to unexpected fingerprint change). Roon will also benefit from being isolated on its own "cage"

You did not specifically mentioned the related background activities but it is clear that disrupts HQPe load distribution and possibly creates extra intracore chatter for HQPe, I suspect you rely a lot on "CPU available time" and Roon makes your algorithm "snap" every time it does a background scanning or other activity.

Now Roon CPU and hardware requirements are fairly low (exception if you have a large library) I understand not everyone has a server with a Hypervisor at home but another small NUC (under $150) fanless small PC could perfectly handle a Roon server separated.

I think for the DSD exotic modulators and even the convolution crowd this would be a very welcome upgrade. Feel free to experiment, I am not going back to a single box. And now I can potentially use HQPOS without needing Roon on it.

 

One more thing and please don't hate, or criticize, I am aware of the possible cons. Roon is running wired on the VM on my home network, the HQPe server is downstairs with my audio system. The HQPe server is connected wireless to the home network, Roon streams 192khz content wireless to the PC downstairs and no glitches. Granted I have a Ruckus 720 AP which cost retail $1300 the AP alone so my wireless home network is not the average one so don't expect to use a cheap AP and get the same results. I can't run a cable downstairs and that is why I was multihomed before and have to rely on Wifi, but I was very skeptical of streaming Roon's output over Wifi, at the same time my wife was watching Netflix and my son was playing an online game on Xbox all over Wifi and a single AP, not a single drop, I was very impressed on how efficient is HQPe receiving the streams from Roon, and I have a questions @Miska, the buffer used on the DAC section is of course on the HQPe output section

Is this buffer proportional or linear to the buffer used between Roon and HQPe? because it seemed very appropriate to me especially over Wifi

 

Again, don't expect to run this over Wifi and get good results and don't expect anyone including @Miska of course to support it, he is been very clear on what's supported and what's not. YMMV

 

 

[luisma]

On 2/4/2021 at 12:21 PM, wolft said:

Hello, which of the new Z590 motherboards from Asus or Gigabyte are recommended for a new music server. Which RAM should be taken? The CPU should be an I9-11900k. Does this CPU make sense. Until then, an I9-10850K will be used.

1. Gigabyte is preferred by Miska, mainly the ones with DACUP USB ports, the extra shielding on these will make it potentially better to connect a DAC. I have done experiments with EMI shielding before and definitely will look for a Gigabyte with this port, this is not available on AMD but it is on INTEL.

 

2. VERY IMPORTANT. Get a board with an INTEL (not REALTEK or other) network card, you will have infinite headaches if you go with Realtek, I know from experience.

 

3. Do your DD (due diligence) on Gigabyte boards, I would not worry too much about the VRM's and such if you are not overclocking, if you are then you need to add getting good and plenty of VRM's to do so.

 

Aside of that you could go with ASUS as well, I just don't know if they do something special with the USB DAC port, last time I checked they did not.

 

IMO any RAM would do, look for low latency (low CAS) ram, doesn't have to be ECC which has a tax on speed over conventional RAM (not much a 2%) but you don't get really any benefits with audio.

 

You haven't specified which music server you are planning on, Windows, Linux, HQP + Roon or just HQP? or something else?

 

[salaryman]

1 hour ago, luisma said:

1. Gigabyte is preferred by Miska, mainly the ones with DACUP USB ports, the extra shielding on these will make it potentially better to connect a DAC. I have done experiments with EMI shielding before and definitely will look for a Gigabyte with this port, this is not available on AMD but it is on INTEL.

 

2. VERY IMPORTANT. Get a board with an INTEL (not REALTEK or other) network card, you will have infinite headaches if you go with Realtek, I know from experience.

 

3. Do your DD (due diligence) on Gigabyte boards, I would not worry too much about the VRM's and such if you are not overclocking, if you are then you need to add getting good and plenty of VRM's to do so.

 

Aside of that you could go with ASUS as well, I just don't know if they do something special with the USB DAC port, last time I checked they did not.

 

IMO any RAM would do, look for low latency (low CAS) ram, doesn't have to be ECC which has a tax on speed over conventional RAM (not much a 2%) but you don't get really any benefits with audio.

 

You haven't specified which music server you are planning on, Windows, Linux, HQP + Roon or just HQP? or something else?

 

I think there’s no more USB-DAC-UP 2 in new Giga motherboard, there’s none in X570/B550 lineup, nor their new Z590s.

 

The last seen likely be the Z490 Vision D. But it will support 11th gen CPU with BIOS update. 

[wolft]

Hello Luisma,

thank you for your post.

I don't see any new Gigybyte Z590 Mainboards with DACUP USB ports. In my setup I want use the following configuration:

Musik-Server --> NAA --> RME Adi-2 DAC Fs --> Lavardin ISx Reference.

My Mainboard favorite is the Asus ROG Strix Z590-F Gaming WIFI. RAM - G.Skill TridentZ RGB Series DDR4 16 GB: 2 x 8 GB DIMM 288-PIN 4000 MHz / PC4-32000 CL18 1.35 V ungepuffert non-ECC

or G.Skill DDR4 32GB PC 4000 CL16 KIT 2x16GB 32GTZR Tri/Z R 32 GB 4.000 MHz 2 16 (F4-4000C16D-32GTZR).

I planning The Music-Server on Windows 10 Pro x64, Linux and Hackintosh (multiboot) with the HQplayer.

My case will be a Lian Li 011 with Seasonic Prime Fanless PX 80+ Platinum modular 500 watt.

Wolft

 

 

 

 

[luisma]

16 minutes ago, wolft said:

Musik-Server --> NAA --> RME Adi-2 DAC Fs --> Lavardin ISx Reference.

In that case just don't worry about the USB port since the NAA will talk to the ADI-2 directly.

Your Asus is fine as it has an Intel LAN port

It also has plenty of slots for adding cards in case you want to do convolution or matrix processing with HQP or just offload the filters

Important the 4 slots for memory which will allow you to add plenty

 

Go for it I would say

[Zauurx]

I have just reached the maximum of the optimization of my H3 with Gigabyte card Z370N (used purchase) + i5-8400.
With the 2 Intel inside ethernet cards (Ultra Durable ™ 25KV ESD and LAN protection against 15KV overloads), I am amazed.

Music Server (Roon eth1 + HQP eth2) > NAA (IPV6) > Allo USBridge > RME > Audia FL100

[Miska]

On 2/7/2021 at 8:09 PM, luisma said:

Is this buffer proportional or linear to the buffer used between Roon and HQPe? because it seemed very appropriate to me especially over Wifi

 

No, it is totally unrelated. That is output side hardware block size, not related to input side.

 

[luisma]

3 hours ago, Miska said:

No, it is totally unrelated.

Thanks, whatever it is at the input it is very very forgiving, or the protocol (is that RAAT?) it is very efficient. I am surprised with the results and even more on how much difference separating Roon does, I always looked at htop utilization of Roon vs HQPlayer and Roon's was almost nothing so you tend to think that Roon will take what? 3%-5% and HQPlayer 95% but these percentages are deceiving as it won't describe the actual scenario, that 5% percent of Roon could mean (with just some sample numbers) affecting HQPe processing for more than 20%. All this is speculation of course, but I am very happy with the results especially running ASDM7EC with no hiccups on a power efficient 5600X on a passive chassis.

Oh and let's no forget that 4.20.1 provides even further optimizations, which is the icing on the cake

[Miska]

15 minutes ago, luisma said:

Thanks, whatever it is at the input it is very very forgiving, or the protocol (is that RAAT?) it is very efficient.

 

Protocols between Roon and HQPlayer are HQPlayer specific.

 

16 minutes ago, luisma said:

I am surprised with the results and even more on how much difference separating Roon does, I always looked at htop utilization of Roon vs HQPlayer and Roon's was almost nothing so you tend to think that Roon will take what? 3%-5% and HQPlayer 95% but these percentages are deceiving as it won't describe the actual scenario, that 5% percent of Roon could mean (with just some sample numbers) affecting HQPe processing for more than 20%.

 

CPU loads don't tell the whole truth because Roon does quite a lot of I/O which doesn't show up there, to some extent only as system load. Also Roon's load is not consistent, but depends what kind of background operations (scanning library, chattering with cloud servers) it decides to do at a particular time, so the result may be different at different times.

 

HQPlayer pretty much assumes to "own" the computer, but usually OS scheduler keeps nasty things out of the way (although Windows is not particularly good on that).

 

[toddrhodes]

I was playing around last night with Sinc-L and just want to see if my theory is correct as to why I can't get it to work for 44.1 or 48K sample rates. Is the "conversion ratio" higher when going from 44.1 --> DSD256 than from say 24/192 --> DSD256? Because at 88.2 and higher, Sinc-L works just fine. But for 44.1 and 48, I never even get playback. I see some CPU and GPU activity for a few seconds, then nothing (and no sound either). It's kinda odd behavior so wanted to double check, For now, I have 1X rates doing either EXT2 or XTR-2S and nX doing Sinc-L and that seems to work great. This is using ASDM7EC modulator for all testing. So if my thinking is correct, the machine has enough horsepower for the fewer "taps" needed when converting higher sampling rates to DSD256 using Sinc-L but the number of taps required for 44/48 are just too much for my machine to tackle. And that's fine, I'm just wanting to make sure I don't have a different issue, since I don't even get audio with dropouts - I just get no audio versus audio with zero dropouts or issues starting at 88.2 sampling rates and higher.

 

Machine specs:

i9900KF at 4.9G

16G 3200 MHz RAM

Win10 dedicated machine

NVidia 1650 GPU

Outputting to Pi4 NAA

Multicore is grayed

Cuda offload is checked (and I am running Convolution but it doesn't seem to change anything if I disable it)

Adaptive Rate Family is checked

[Miska]

13 minutes ago, toddrhodes said:

I was playing around last night with Sinc-L and just want to see if my theory is correct as to why I can't get it to work for 44.1 or 48K sample rates. Is the "conversion ratio" higher when going from 44.1 --> DSD256 than from say 24/192 --> DSD256?

 

(44100 x 256) / 44100 = 256

(48000 x 256) / 192000 = 64

 

So yes, former case has 4x higher conversion ratio.