Design a PC/Server for ROON and HQ Player

[Miska]

There is the power side also, of course.

 

Yes, don't put everything on the same wall socket and without proper treatment...

 

In listening room, I have two power cleaners, one for the digital equipment (computers, etc) and one for the analog equipment.

 

DACs are connected to either one depending on whether they run on SMPS or linear PSU. SMPS go to the computer side and linears go to the analog side.

 

Office where the bigger computers are has it's own power cleaners and UPS.

[Miska]

Considering a switch from MAC to PC (only for audio) after 15 years. So now I have to remember how to build a PC. Roon will let me move my core, but I guess I will need to buy HQ Player again once this is done. Sigh.

 

There's a discount for the second platform license, send me an email when you need it...

[Miska]

So I'm trying to get to DSD512 with the poly-sinc-mp filters and my current system does -2s well but hiccups at poly-sinc-mp. Not sure that its a pure CPU issue because my system loads to 32% (6x 3.4Ghz 4930k processor). The GTX 760 doesn't seem to help.

 

Please check out the per-core load graphs in Resource Monitor. If any one of the cores is maxed out, then hiccups are likely to occur. Also note that for CPUs with HyperThreading, the CPU is practically fully loaded at 50% total load, because the virtual second CPU's from HT don't really help much with HQPlayer type of work loads.

 

adding a GTX 1080 I'm assuming this will do DSD512 poly-sinc-mp fine ...

any idea about DSD1024? ...just because ...

 

If the GTX 1080 performance is going to be anything like they promise, that won't be a bottleneck for DSD1024... Bottleneck is more likely to be the parts that still run on the main CPU. Let's see how much faster I can still make those ones.

 

 

Generally, if you go for CUDA, you can use CPU with higher base clock speed and less cores.

[Miska]

Good post. I know I keep reading an i7 is min. for DSD512, but here's a screenshot of my cores while upsampling DSD64 to 256. So do you think for sure that my Mini wont do DSD512 on a Windows bootcamped?

 

I'm pretty sure it won't. Going DSD256 -> DSD512 at least doubles the load, and you have two cores already at around 75%. Based on your load figures, you are now at optimal max of your machine. There's some margin left for the Roon GUI.

[Miska]

My question to the guys who already got DSD512 working: are you able to upsample to 512 and use matrix room correction at the same time without CUDA offload?

 

Depends on your filter lengths and source sampling rate. With RedBook content the extra load of room correction is small, but doing it for DSD sources can get pretty heavy...

[Miska]

It is better to go directly to 6700K or similar Xeon (not familiar to me) , I think it is best to get max GHz cpu if you want to use Roon with HQPlayer and DSD 512 conversion. I am going to install a new PC with 6700K. Water cooled etc. Cost me @USD 1100.00 here.

 

There's no Skylake-based Xeon E5 yet (E5 v5). There is only E3 v5 so far, but it doesn't offer any practical advantage over the i7 6xxx. Advantages of E5 are twice the RAM bandwidth due to four channel memory controller and support for multiple sockets on bigger models. Plus support for ECC RAM which matters to me for reliability reasons (computer running 24/7).

[Miska]

Now Nvidia is starting to roll out their newest Pascal-generation GPUs on the Quadro-series:

New Quadro Graphics Built on Pascal Architecture | NVIDIA

 

Let's see how the price/performance fares out...

[Miska]

Is there any need to include a graphics card in a new PC for HQPlayer? It would make the build cheaper to use just the Intel i7 6700K which will do all DSD/PCM upsampling to DSD512 (cheaper case, cooling and PSU).

 

Depends on what you want to do exactly. If you need digital room correction, or matrix processing, you can get quite a bit of extra boost from a GPU.

 

According to https://developer.nvidia.com/cuda-gpus the NVIDIA TITAN X, GeForce GTX 1080, GeForce GTX 1070, GeForce GTX 1060 have Compute Capability 6.1 and GeForce GTX 980 Ti, GeForce GTX 980 (Jussi), GeForce GTX 970, GeForce GTX 960, GeForce GTX 950 Compute Capability 5.2 So it looks like even the latest cards don't give more than +20% performance but there's a big price difference!!

 

Compute capability doesn't directly translate to performance, it is just an architecture version number indicating certain feature set.

 

In addition to pure performance, the newer Pascal generation (1000-series) gives better power/performance ratio because it uses smaller manufacturing process and lower voltage.

 

If I did want to upgrade to DSD1024 or something else in the future, I'd need a new DAC and probably a new PC so no need to build for graphics upgrade capability? Or would you recommend building the PC capable of taking a CUDA GPU which would enable the 6700K to cope with DSD1024? "Poor man can't afford to buy cheap!"

 

I would recommend to get started with just 6700K, but get a case big enough so it can house at least GTX 1060. You can then add GPU later and swap PSU if necessary.

[Miska]

@Miska, for ultimate performance, say i7-6900k with watercool/overclock or i7-7700k? ... along with Pascal

 

I don't know anything yet about 7700K. But it depends on how many channels / what kind of processing you need.

 

My new server is 6950X (Broadwell-E) for doing 8-channel output to exaSound e28. I posted some initial testing results on the FB page.

[Miska]

Yes, I'd like to be able to do DSD1024 along with kernel if possible and the question will be either LGA 1151 or 2011-3

 

Look for something that has as good as possible per-core performance, but still with enough cores.

 

Main difference between 1151 and 2011-3 is memory bandwidth and PCI-Express lanes. 1151 has two memory channels while 2011-3 has four. 6700K has 16 PCI-Express lanes while 6900K has 40. However, 6700K has highest per-core speed.

[Miska]

What do you think of this one EVGA 04G-P4-2983-KR GeForce GTX 980 SC Gaming ACX 2.0 Graphics Card? Seems good value at £280. Also found this indicating GTX 980 SLI is the best performing over GTX 1060/1070 and the other 9xx cards:

[ATTACH=CONFIG]29542[/ATTACH]

 

That table seems to refer to mobile (laptop) variants.

 

HQPlayer cannot utilize SLI setups (multi-GPU) at the moment. From HQPlayer perspective, only meaningful columns are the two double precision floating point ones. GTX 980 is close GTX 1060. I have tested ASUS Strix-series GTX 980 and GTX 1060 (with 6 GB RAM) and the GTX 1060 was slightly faster.

 

3GB model of GTX 1060 is around 250€.

 

 

Depending on your computer setup, you may want to look for a card that is as silent as possible. Another important aspect to note is size of the card. For example the ASUS Strix-series 1060/1070/1080 are HUGE and may be hard to fit.

 

For my Linux desktop I got the ASUS Dual-series 1060 card because it is smaller:

DUAL-GTX1060-O3G | Graphics Cards | ASUS Global

There are also 6 GB 1060 and 8 GB 1070 models in the Dual-series. Also feels cheaper than the Strix-series though.

[Miska]

Now there's a blog post about my most recent server build:

http://www.computeraudiophile.com/blogs/miska/building-new-server-multichannel-processing-702/

[Miska]

That link isn't working

 

Now it should work. It was in draft status so displayed only to me. It wasn't immediately apparent to me, now I changed status to "publish now".

[Miska]

Have you guys just compared your 4GHz cpus and gamer GPU setups to a Melco? Don't get me wrong, I have Haswell-E and GTX 980 SLI for gaming, but everytime I upgrade my rig my sound seems gets worse. Adding a second graphics card was the worst.

 

If you feel that adding more computational capability makes sound worse, try adding a NAA. It'll move out the actual audio output to another low power device. Sonore's microRendu, SOtM SMS-200 and exaSound PlayPoint being examples of such. This way you can have both, plenty of power to do all the fancy DSP and low noise device for output.

 

However, I feel that the machine I built is pretty good for the purpose alone, I have the exaSound e28 directly connected to the DAC-UP USB port and the sound is good. One thing to note is to make sure you have a good PSU. Two cards in SLI setup can alone draw 400W, so to be on the safe side, you should have 800+W PSU.

 

I also have a fanless small CL100 to act as a Windows-based NAA with T+A DAC8 DSD running at DSD512. But as a NAA it could also handle the exaSound e28. I replaced CL100's poor stock wallwart with a better PSU.

[Miska]

What size ssd and RAM do you have in the CL100 NAA?

 

RAM is the max 16 GB, although it would probably be just fine with 8 (as long as it doesn't use swap). SSD is 256 GB split into two 128 GB partitions, one fro 64-bit Win10 Pro and another partition for Linux.

[Miska]

RAM is the max 16 GB, although it would probably be just fine with 8 (as long as it doesn't use swap). SSD is 256 GB split into two 128 GB partitions, one fro 64-bit Win10 Pro and another partition for Linux.

 

Oh, and I took it without any wireless options, on purpose. I use only the gigabit ethernet and wanted to avoid any possibly RFI from wireless.

[Miska]

In your opinion, is it overkill for hq players's most demanding stereo upsampling?

 

Yes, I'd say it's an overkill for stereo to two channel output cases. If extra power is needed for such cases, CUDA offload may be better approach, for example with GTX 1060.

 

Even for stereo source content, one could still want to use this kind of machine to output multichannel with digital cross-over filters or something similar. Or alternatively perform something like stereo mix-down of multichannel DSD256. But there are also those 6 and 8 core versions in the Broadwell-E series (such as 6850K and 6900K).

 

And is it too noise to be kept near the hifi system, when being driven hard?

 

I have not yet really pushed the machine to the edge. I have ordered bunch of quiet Noctua fans I will replace to the case and also connect one of the 12 cm fans to be controlled by the GPU instead of motherboard. So some fine tuning still to happen.

 

Now it's not bad if placed properly.

[Miska]

I have a GeForce 1080 Founders Edition card in my new build. Ever since, I've not been able to use Roon b/c I get OpenGl 3.0 isn't installed. But I have the latest driver from nVidia.

Has anyone had similar problems...solutions?

 

The latest drivers from nVidia should work fine, up to OpenGL 4.5. I don't think the drivers make difference between 1080 Founders Edition and the other versions, since the GPU itself is same. Possibly some driver installation issue? You could try to uninstall all nVidia items through Windows Control Panel, reboot and then reinstall the driver package.

 

I didn't play with Roon yet on the new machine, but at least HQPlayer's use of OpenGL is functional (splash-screen at startup and touch/fullscreen mode). Sometimes OpenGL can be tricky, so there could be sometimes some compatibility issues when certain OpenGL extensions are used by the application.

[Miska]

What is the CPU usage like with no CUDA off loading same music "5.1 channel 48/16 DVD-rip, upsampled to 12.3 MHz DSD256 using full poly-sinc-short-mp and ASDM7"?

 

[Miska]

Yes, but what is with the very high levels of Internet usage? I have noticed this as well when using the Roon/HQPlayer combination. Is that just the system confusing the bitrate toward the NAA as "internet traffic" or is one of HQP/Roon actually making meaningful contact with the cloud?

 

In the cases I posted there's no Roon and no NAA in use. But HQPlayer is playing content from my local file server because the server doesn't have any other storage than the 512 GB M.2 SSD for the OS. Since the content is FLAC, the overall network usage is somewhere around 1 Mbps or so.

 

But the machine has good network interface with offload capabilities. If the output would be sent to NAA, there would be 93.75 Mbps stream. Which is 10% of the 1 Gbps network capacity, so essentially still almost nothing.

[Miska]

Is CUDA offloading operational in the latest 3.14 production? I built my machine to your specs back in March

http://www.computeraudiophile.com/f11-software/design-pc-server-roon-and-hq-player-25796/index6.html#post536775

and hope my GPU card

(Asus GeForce GTX 960 Overclocked 4GB GDDR5 PCI-E Video Card)

is not obsolete after a few months now.

 

Yes it is, and the GTX 960 shouldn't be obsolete anytime soon.

[Miska]

OK please keep us posted of your findings. I for one can't afford to experiment with £k components until I get the right mix The two things I've read are important are RFI and jitter. If both PC's were connected by optically isolated Ethernet that just leaves jitter, unless RFI was getting into the bad sounding PC via signal or power cables.

 

Ethernet doesn't have jitter given the way how NAA works... Copper ethernet is transformer isolated, but optical ethernet of course isolates even one step better.

[Miska]

Also, a Ravenna card inside Merging DAC is a kind of NAA, isn't it?!

 

At conceptual level yes, but otherwise quite different. If/when the DAC is not timing master in Ravenna, then it is subject to jitter, because clock is distributed through PTP. So it is different from NAA in that respect. Overall, the Ravenna protocol is quite different, since it was designed for entirely different purpose and audience.

 

In the same way Merging has now implemented Roon Server onboard, they could implement Sygnalist's nad if they wanted to.

 

Yes, I think so.

[Miska]

I am looking at Video cards for Cuda offloading, the 1060 series so is 3GB sufficient for Cuda processing or should I only be looking at the 6Gb size?

 

If you don't do convolution stuff, the RAM size is not going to be an issue with 1000-series.

 

Main item to consider is how much processing power you need and select the GPU according to that. If your DAC supports also 48k-base DSD rates (so you can use "Auto rate family"), you should be fine with 1060 up to DSD512. If it doesn't, at least the 1060 one I have in one of my machines cannot do 48k -> 22.5792M with heaviest poly-sinc. I didn't try yet the same with 1080.

 

Could be that I just need to try to do some more optimizations and then 1060 would be sufficient too. At the moment I feel the memory bandwidth becomes problem more than raw power.

[Miska]

@ Miska Thank you again, My personal dac (heavily modified Gustard X20u) I don't believe supports 48K digital, the T&A I am building a system around for a friend, of course does, so that is good news for him. Curious is there some CPU load chart showing which filters are most intensive, maybe from most intensive to least? For some reason I can usually do Poly sinc ext with the 4790K just fine, The regular Poly Sinc's stutter and any (non 2s) poly sinc MP won't even play. I assume closed form is the most intensive, with MP (non 2S) next and then .......?

 

-mp versions are equal in runtime load to the non-mp versions, only difference is that -mp versions have longer initialization time, so starting playback will take somewhat longer.

 

-ext is roughly equal load to "poly-sinc" but it has more restrictive ratio rule set. So if -ext works for you, you would get more even load if you just check the "Auto rate family" in settings.

 

-hb is heaviest and then poly-sinc/poly-sinc-mp next.

 

closed-form has different load structure and can utilize more CPU cores, so there is no simple rule to compare it against other filters.

 

All -2s versions are significantly lighter than non-2s, from simplified technical point of view the quality is equal between the two. But the two are not exactly the same, so the calculated sample values are not exactly the same within the processing resolution range.