Design a PC/Server for ROON and HQ Player

[Miska]

Is HQPlayer still CPU bound whilst setting up for upsampling, i.e. a more powerful GPU wouldn't reduce cold startup times?

 

Yes, the initialization supports multi-core CPUs, but doesn't support GPU offload yet. That is still on my TODO-list...

 

I just did some further optimizations for initialization of the convolution engine.

[Miska]

That damn Quadro M4000 of yours is a $900 card! Yikes.

 

Isn't that still cheap compared to most high-end audio electronics? And it has at least billion times more transistors than most audio electronics! ;D

 

I decided to go with GeForce GTX 980 for the Win10 machine because people were asking for less expensive alternative. See here:

Upgrade GPU for more CUDA processing power - Blogs - Computer Audiophile

 

I use Quadro on my Xeon Linux-workstation.

 

What do you think is an inexpensive card that will help me when CUDA offloading multichannel? Also, I am going i7 6700k, a GPU card maybe like this https://pcpartpicker.com/part/evga-video-card-04gp41962kr

and a fiber card. You think 520w is still fine? PCpartspicker has me at 314W before the fiber card.

 

Depends on what you want to do exactly. For multichannel you'll need quite a bit of processing power and RAM on the card. GTX 960 is probably minimum for that kind of work (about 50% of speed of GTX 980).

 

520W should be fine for it, nVidia recommends 400W when used with 3.2 GHz i7:

GeForce GTX 960 | Specifications | GeForce

 

I chose the card based on quiet operation, so it has large heatsink with two slowly rotating fans to keep acoustics down. And the GTX980 model just fits in the case and 520W. The case can accommodate full length cards by removing one block of 3.5" drive bays, but putting bigger PSU would mean having a PSU with cooling fan.

[Sagittarius]

Isn't that still cheap compared to most high-end audio electronics? And it has at least billion times more transistors than most audio electronics! ;D

 

I decided to go with GeForce GTX 980 for the Win10 machine because people were asking for less expensive alternative. See here:

Upgrade GPU for more CUDA processing power - Blogs - Computer Audiophile

 

I use Quadro on my Xeon Linux-workstation.

 

 

Just a quick question, does HQ player support acceleration via Intel integrated GPUs?.

[ismewor]

Miska,

I would like to know does CUDA will support GTX970m with 3G memory.

 

Thx

Adrian

[Miska]

I would like to know does CUDA will support GTX970m with 3G memory.

 

All the 900-series should be fine at varying amounts of processing power.

 

Soon nVidia should be launching a new 1000-series based on their Pascal architecture that should give a huge performance boost for HQPlayer (64-bit double precision floating point). There's some more info here.

[kravi4ka]

Any idea about the price range for that? If it is in GTX980 Strix category we'd better wait indeed but if it is in the five figures...

[gldgate]

Any idea about the price range for that? If it is in GTX980 Strix category we'd better wait indeed but if it is in the five figures...

 

Pascal graphics card to launch at Computex 2016 and enter mass shipments in July

 

The usual roll-out is to start with high end/high $ "enthusiast" cards to get the press and consumers jacked up and follow with more mainstream offerings a few months later. I'd bet replacement for 970 and 980 cards will start showing up in September/October time frame.

[ismewor]

Thanks Miska,

 

So with that said. Does AMD totally out of the picture for the DFP method?

 

Thanks,

Adrian

[Miska]

So with that said. Does AMD totally out of the picture for the DFP method?

 

No, it only depends when their GPUs are programmable in some suitable way. OpenCL is not among the suitable methods.

[sydlow]

Just finished a build. Thought I'd share:

- Gigabyte GA-Z170MX-Gaming 5 selected because it's the smallest mobo with "USB DAC-UP" feature promising lower noise from PC to DAC on two USB ports. You can also turn off power to these ports if your DAC doesn't need it.

- Intel 6700

- 16GB RAM

- Streacom case

- Streacom Zeroflex PSU (internal)

 

It's running Roon Server and HQP just fine into a Auralic Vega as DSD128.

CPU cores running at under 15% load.

Very sweet box. Gorgeous sound coming out of the Vega.

[Solstice380]

RAM Question:

 

For loading in 32 GB of RAM, is it better to use 4x8 or 2x16?

[sdolezalek]

All the 900-series should be fine at varying amounts of processing power.

 

Soon nVidia should be launching a new 1000-series based on their Pascal architecture that should give a huge performance boost for HQPlayer (64-bit double precision floating point). There's some more info here.

 

If the focus is on audio processing, is there any reason not to go in the direction of one of these: NVIDIA Tesla K40C 12GB Computational Accelerator (NVIDIA Tesla K40C 12GB Computational Accelerator(753960-B21)| Hewlett Packard Enterprise® ) rather than a new graphics card?

[Miska]

If the focus is on audio processing, is there any reason not to go in the direction of one of these: NVIDIA Tesla K40C 12GB Computational Accelerator (NVIDIA Tesla K40C 12GB Computational Accelerator(753960-B21)| Hewlett Packard Enterprise® ) rather than a new graphics card?

 

No, just price and generation. K40C is a bit older Kepler-generation and current equivalent is Maxwell-generation M40. nVidia just released a new top of the line Tesla P100 with their newest Pascal architecture product. But there is no information yet on when P40 product would materialize...

 

The launch presentation gives some kind of picture:

[Miska]

RAM Question:

 

For loading in 32 GB of RAM, is it better to use 4x8 or 2x16?

 

For best performance it is best to have all memory channels in use. So check what kind of memory bank designations your system has.

[Solstice380]

For best performance it is best to have all memory channels in use. So check what kind of memory bank designations your system has.

 

Tnx. 4x8 it will be! [emoji12]

[kravi4ka]

What about the speed of the RAM, is 2133 OK and is there a point to go higher?2400,2666? TIA

[EuroDriver]

RAM Question:

 

For loading in 32 GB of RAM, is it better to use 4x8 or 2x16?

 

With RAM, what we have found is less RAM is more sound quality.

 

4 memory channels in use gives better computing through put performance, but guess what using 2 memory channels can sound noticeably better. It's not huge, but if you want to squeeze the last drop of SQ ;-)

[EuroDriver]

What about the speed of the RAM, is 2133 OK and is there a point to go higher?2400,2666? TIA

 

RAM specs, and then adjusting voltages and timings to get the best sound. It makes re-stringing a Stradivarius look simple by comparison !

[kravi4ka]

Thanks! As I said in the other thread - there are not many people ready to spend the time and money to try the million variables in practice, I have greatest respect for what you are trying to do and honeslty wish that the world reads a bit more about your server. Is there a way you can talk to a computer manufacturer to do a custom board for you?

[dean70]

I would imagine DDR3L Ram on Skylake processors would lead to a further improvment.

[Miska]

I would imagine DDR3L Ram on Skylake processors would lead to a further improvment.

 

Skylake is intended for DDR4. Look out for lowest possible CL (CAS latency) to maximize memory speed.

 

I'm got Kingston HyperX FURY kit of 4 DIMMs:

FURY DDR4 Memory | HyperX

 

For even higher clock speeds they have the Predator series:

HyperX Predator DDR4 Memory for Intel X99 Motherboards

[Solstice380]

I'm looking at the Kingston Savage with a CL of 12.

 

HX424C12SB2K

 

But, it takes 1.35V vs 1.2 for the Fury. I can get it in the US for only US$10 more than Fury. Any downside other than voltage and heat? Maybe lifetime?

[Solstice380]

Plus the Savage is XMP.

 

And, oops, it's US$20 more at 4x8G.

[Miska]

I'm looking at the Kingston Savage with a CL of 12.

 

HX424C12SB2K

 

But, it takes 1.35V vs 1.2 for the Fury. I can get it in the US for only US$10 more than Fury. Any downside other than voltage and heat? Maybe lifetime?

 

Looks good to me, the Fury is CL14 so if you get CL12 and higher clock speed, then it should be quite a bit faster.

 

Lifetime shouldn't be affected as long as there's sufficient cooling.

[Solstice380]

I fell into a good buy from a friend on a build he had made for him. I caught him at a weak moment when he needed cash!

 

Asus Z170 Deluxe mobo, i7-6700K, water cooled, Asus 970x 4Gb graphics, Intel USB 3.1, 8Gb SDDR4 G Skill 2400, 720W Seasonic PSU, Win 10 Pro, in a nice case. Now looking to increase RAM. The processor is running at 4.16 MHz right now and the fan for the CPU water cooler never goes above idle. There is a fan on the graphics card and 2 - 120mm case fans at 600 RPM idle, 1 in and 1 outflow.

 

Stripped down Win install, pretty clean. Still running too many processes but I'd say it sounds about as good as my Uptoned Mac Mini. Sure like seeing Closed-form filter doing 44/16 to DSD128 (DoP) only pushing the CPU to 15%.

 

My goal is to use this for a Roon/HQPlayer server so I'm interested in more RAM. It has 2 network ports so I'm going to try setting up a direct network connection between the two like Alex C. did, with the Mini as a Win/ASIO/NAA, so I can keep it off my main network. If it is problematic I'll go the router route.