Best Nvidia CUDA Card for HQPlayer

[Miska]

48 minutes ago, Andrey Boldakov said:

Miska, could you kindly answer if multi-GPU (dual GPU) works for offloading HQPlayer tasks? 

 

I'm not sure. I never did any notable testing on such setup. And Nvidia doesn't document how CUDA operates for example in SLI setup.

 

At the moment HQPlayer doesn't have any balancing between GPUs on it's own, outside of CUDA's inner workings. I'm not sure if CUDA always gives me the same GPU.

 

[Andrey Boldakov]

2 hours ago, Miska said:

I'm not sure. I never did any notable testing on such setup.

Ok, just DSD128 requires 4Gb of memory

DSD265 will ask for 8Gb

DSD 512 for 16Gb

But if I wanna give a try to i9-12900KS for DSD1024 so I need GPU with 32Gb of video memory what is hard to achieve within one GPU (((

(thinking over Holo May and DSD1024 EC codec if it's even possible on that planet)

[Miska]

1 minute ago, Andrey Boldakov said:

Ok, just DSD128 requires 4Gb of memory

DSD265 will ask for 8Gb

DSD 512 for 16Gb

But if I wanna give a try to i9-12900KS for DSD1024 so I need GPU with 32Gb of video memory what is hard to achieve within one GPU (((

(thinking over Holo May and DSD1024 EC codec if it's even possible on that planet)

 

Depending on GPU generation, it can go like 3/6/12/24 GB RAM.

 

But that applies only to sinc-L filters. If you use something else, amount of RAM is not such an issue.

 

[Andrey Boldakov]

13 hours ago, Miska said:

But that applies only to sinc-L filters. If you use something else, amount of RAM is not such an issue.

Wow, that's very important info! 

Yes, I've migrated from Sync-L to gauss-xla (some good man here suggested me, and I'm really happy with gauss-xla in conjunction with ASDM7ECv2) 

 

So, if I'm trying to reach DSD1024 with i9-12900k, so what is the lowest level of GPU can really help? Maybe some ancient one like Nvidia Tesla K80 24gb DDR5, or some newer but low end like Nvidia Quadro RTX A2000? 

 

Is it "must have" using top of the line GPU to help exactly in DSD1024 with EC modulators and gauss-xla? 

 

The main pain I need to avoid is buying GPU that will help in nothing (I mean will not help at all) with DSD1024 ASDM7ECv2 gauss-xla

[Miska]

8 hours ago, Andrey Boldakov said:

Is it "must have" using top of the line GPU to help exactly in DSD1024 with EC modulators and gauss-xla? 

 

You will need pretty hefty GPU to do DSD1024 with gauss-xla. I don't have exact spec of what would be suitable.

 

[Andrey Boldakov]

45 minutes ago, Miska said:

You will need pretty hefty GPU to do DSD1024 with gauss-xla. I don't have exact spec of what would be suitable.

Hmmm... Does it mean the processing can't be shared between CPU and GPU? Does GPU must do the job with gauss-xla completely? 

I mean i9-12900K can almost do the DSD1024 alone, and it woukd be great to help a little by GPU... But does it work in another way where GPU can not be used as CPU-helper? 

[Miska]

30 minutes ago, Andrey Boldakov said:

Hmmm... Does it mean the processing can't be shared between CPU and GPU? Does GPU must do the job with gauss-xla completely?

 

You can choose if you want to offload filter and/or convolution work to the GPU. When offload is active, that portion of the work is executed on GPU.

 

30 minutes ago, Andrey Boldakov said:

I mean i9-12900K can almost do the DSD1024 alone, and it woukd be great to help a little by GPU... But does it work in another way where GPU can not be used as CPU-helper? 

 

This is quite complex topic. One aspect of GPU offload is that it reduces load on number of CPU cores, allowing higher turbo boost clocks for the cores that handle tasks that can be executed only on CPU.

 

[Andrey Boldakov]

7 hours ago, Miska said:

allowing higher turbo boost clocks for the cores that handle tasks that can be executed only on CPU.

That says i9-12900KS may perform better than i9-12900K with GPU.. at least in terms of value for money

[shoubhikbhatti]

Hi,

I want to use HQPlayer to upsample to DSD 512/104. I have i7 12700K CPU and EVGA RTX 3060. Would it be enough? Also for confirmation the HQPLayer official site says Nvidia GPU, I am assuming that any GPU with Nvidia Architecture (with base cuda support) and not only GPUs actually manufactured by Nvidia?

 

[AudioDoctor]

2 hours ago, shoubhikbhatti said:

Hi,

I want to use HQPlayer to upsample to DSD 512/104. I have i7 12700K CPU and EVGA RTX 3060. Would it be enough? Also for confirmation the HQPLayer official site says Nvidia GPU, I am assuming that any GPU with Nvidia Architecture (with base cuda support) and not only GPUs actually manufactured by Nvidia?

 

 

Sure it will work. Will it work to the level you want it to work?  No idea. Try it.

 

Any card on Nvidia architecture will do it, as far as I know, yes.

[Miska]

16 hours ago, shoubhikbhatti said:

Also for confirmation the HQPLayer official site says Nvidia GPU, I am assuming that any GPU with Nvidia Architecture (with base cuda support) and not only GPUs actually manufactured by Nvidia?

 

Nvidia manufactures the GPU chips. Then various graphics card and laptop manufacturers put those chips on their products. As long as product has a compliant Nvidia GPU chip, it should work.

 

Nvidia, AMD and Intel manufacture very few end products, they are primarily just chip makers.

 

[Cornell77]

On 4/23/2022 at 7:54 PM, Miska said:

 

Depending on GPU generation, it can go like 3/6/12/24 GB RAM.

 

But that applies only to sinc-L filters. If you use something else, amount of RAM is not such an issue.

 

I would like to choose a GPU that can help up-sample to DSD512 using sinc-L.  For the RTX3000 series, is 12GB sufficient? I recognize that many factors are involved as well.  I would like to make sure that I understand the minimum need.  In the conversation, it was mentioned that 16GB may also be required  for sinc-L. depending upon the GPU generation.

[AudioDoctor]

5 hours ago, Cornell77 said:

I would like to choose a GPU that can help up-sample to DSD512 using sinc-L.  For the RTX3000 series, is 12GB sufficient? I recognize that many factors are involved as well.  I would like to make sure that I understand the minimum need.  In the conversation, it was mentioned that 16GB may also be required  for sinc-L. depending upon the GPU generation.

 

For that you're going to need to go big and get a 3090 and the most powerful CPU you can get for HQPlayer.

[pis99]

5 hours ago, Cornell77 said:

I would like to choose a GPU that can help up-sample to DSD512 using sinc-L.  For the RTX3000 series, is 12GB sufficient? I recognize that many factors are involved as well.  I would like to make sure that I understand the minimum need.  In the conversation, it was mentioned that 16GB may also be required  for sinc-L. depending upon the GPU generation.

No, 512DSD 7EC module will not do Sinc-L at base rate but it will do 88.2/96 files if your have 3090 with 24G. With 12900KS + RTX A6000/48G 1024DSD/Sinc-L is achievable at DSD5EC and below. Sinc-L is a different type of filter that requires a lot of things , even CUDA versions matter.

 

 

[Rovo]

Miska advised me to use a NVidia GPU with minimum 16GB. I am using now an Intel i9-10900K with RTX A4000 16GB. With this combination I can upsample everything sinc-L + AMSDM7EC 512+fs with a base rate of 44,1kHz to DSD512 and with a base rate of 48kHZ to DSD256. 

[Cornell77]

8 hours ago, Rovo said:

Miska advised me to use a NVidia GPU with minimum 16GB. I am using now an Intel i9-10900K with RTX A4000 16GB. With this combination I can upsample everything sinc-L + AMSDM7EC 512+fs with a base rate of 44,1kHz to DSD512 and with a base rate of 48kHZ to DSD256. 

Thank you very much for this input.  I am now considering purchasing the RTX A4000 video card.  How loud is the A4000 fan at idle?  How loud does it get when under the load of DSD512 processing?  Have you ever measured the GPU temperature?

 

[Rovo]

In my setup the RTX A4000 fan is running at approximately 40 to 50%. The GPU temperature is approximately 70 degrees Celsius. This information is available through the NVidia tool. I have the impression that the GPU temperature is kept at approximately 70 degrees and the fan speed is adjusted accordingly. At the moment I have also some case fans and CPU fans running. At a distance of 1 meter I can hear the fans a little without it disturbing me. The fan of the GPU is only a small part of this noise. I would expect that I would not hear the GPU fan at 1 meter if this was the only fan running. The computer case is a Fractal Design Define 7. 

[Cornell77]

I recently purchased an RTX A4000 video card to pair with my I7-11700k CPU.  I am now able to upsample Qobuz files to DSD 512 using ASDM7-512 fs as well as ASDM7ECv2 modulators with filters such as poly-sinc-ext3 or poly-sinc gauss-xla.  CPU utilization  runs just over 40% and GPU utilization also maxes out at  40%.  (Note that GPU utilization is measured by Asus GPU Tweak 3 software as Windows task manager yields incorrect GPU measurements.) Total system power consumption remains comfortably under the 450 watt power supply used for the build.  Needless to say, I am pleased with my choice and would like to thank Rovo for this recommendation.

 

Here is a chart that shows a number of graphics cards, both classic and current, and some relevant specifications:

 

Graphics Card    FP64 performance    Memory  Size (GB)    Memory Bandwidth (GB/s)

GTX 1060                       137                             6                              192
RTX 3060                       199                            12                             360
Quadro P5000               277                            16                             288
RTX 3070                       317                              8                             448
RTX 3080                       465                            10                             760
RTX 3090                       556                            24                             936
RTX A4000                     599                            16                            448
RTX 3090ti                     625                            24                           1008
RTX A5000                     868                            24                             768

 

Explanation:

 

HQ Player uses 64 bit floating precision to perform its computationally intensive task of up-sampling.  The reason why this is important is that Nvidea’s pipeline for its Pro cards typically offers twice the bandwidth for FP64 processing in comparison to its consumer graphics cards. This 2X multiplier is a conscious decision on the part of NVidea to differentiate their pro line from their consumer line of graphics cards.  This Pro series advantage can make a noticeable difference in the power needed to run HQ Player effectively.  

If you look at the above table, you can see the strong performance of the RTX A4000 graphics card.  For playing games, the RTX A4000 is considered to be the rough equivalent to an RTX 3070.  Reviews confirm this performance bracket for gaming.  However, for HQ Player, the RTX A4000 slots in closer to the RTX 3090/RTX 3090ti.  Total memory available is also important if you plan on upsampling to DSD 512 with demanding filters such as sinc-L.

 

I hope this information proves useful to others in making their choice of GPU.  Finally, as Miska points out, CPU power is needed for the modulator, so your CPU and GPU should reflect a balanced choice depending upon your goals.

 

[AudioFanatics]

@Cornell77 May you help me measure which exactly gpu memory & cpu usage for convert 44.1 Khz and 48 kHz to DSD256 with sinc-L & ASDM7ECv2 in your set up? I'm planning to build a PC for HQPlayer and now very confused about GPU choice. Thanks in advance.

[Rovo]

GPU memory usage is a little above 6 GB. CPU usage approximately 40% (from memory) for a i9-10900K. 

 

For me the choice for a RTX A4000 was determined by the required power of the GPU. Wanted to have it as low as possible due to the ULPS I wanted to use. 

[Cornell77]

6 hours ago, AudioFanatics said:

@Cornell77 May you help me measure which exactly gpu memory & cpu usage for convert 44.1 Khz and 48 kHz to DSD256 with sinc-L & ASDM7ECv2 in your set up? I'm planning to build a PC for HQPlayer and now very confused about GPU choice. Thanks in advance.

The need to upsample to DSD256 (as opposed to DSD512) makes your task easier.  I was able to upsample to DSD256 with poly filters using only the I7-11700k without GPU.  Doubling the upsampling rate to DSD512 will effectively double the total compute power required.

 

I am personally not a fan of the sinc-L filter.  It is an integer filter, which means that to upsample 48kHz PCM to DSD,, your DAC must handle 48x DSD.  Most DACs including mine do not have this capability.  When using sinc-L to upsample PCM to higher rate PCM, I found it to be an analytical sounding filter that presents great precision in the reproduction of detail - but is leaner in presentation than I'd like.

 

If I was to build a new PC, I would wait a few months for the release of Intel's new Meteor Lake CPU's and matching Z790 motherboards.  An Intel I7-13700k is probably all that you need (without a GPU) for DSD256 upsampling.  This chip may very well handle DSD512 by itself.  Having some margin with your selection will be a good thing for the future. You can find early benchmarks for the I7-13700k by using a search engine.  If I needed to add a GPU, I can't speak highly enough for the RTX-A4000.  It draws very little power so that my total system power consumption hovers around 300 watts.  The small fan and single slot size makes for a hot running card with slightly higher noise than would be ideal, but overall it is a great choice for HQ Player.

 

Hope this helps.

[shoubhikbhatti]

On 7/17/2022 at 12:59 PM, Cornell77 said:

I recently purchased an RTX A4000 video card to pair with my I7-11700k CPU.  I am now able to upsample Qobuz files to DSD 512 using ASDM7-512 fs as well as ASDM7ECv2 modulators with filters such as poly-sinc-ext3 or poly-sinc gauss-xla.  CPU utilization  runs just over 40% and GPU utilization also maxes out at  40%.  (Note that GPU utilization is measured by Asus GPU Tweak 3 software as Windows task manager yields incorrect GPU measurements.) Total system power consumption remains comfortably under the 450 watt power supply used for the build.  Needless to say, I am pleased with my choice and would like to thank Rovo for this recommendation.

 

Here is a chart that shows a number of graphics cards, both classic and current, and some relevant specifications:

 

Graphics Card    FP64 performance    Memory  Size (GB)    Memory Bandwidth (GB/s)

GTX 1060                       137                             6                              192
RTX 3060                       199                            12                             360
Quadro P5000               277                            16                             288
RTX 3070                       317                              8                             448
RTX 3080                       465                            10                             760
RTX 3090                       556                            24                             936
RTX A4000                     599                            16                            448
RTX 3090ti                     625                            24                           1008
RTX A5000                     868                            24                             768

 

Explanation:

 

HQ Player uses 64 bit floating precision to perform its computationally intensive task of up-sampling.  The reason why this is important is that Nvidea’s pipeline for its Pro cards typically offers twice the bandwidth for FP64 processing in comparison to its consumer graphics cards. This 2X multiplier is a conscious decision on the part of NVidea to differentiate their pro line from their consumer line of graphics cards.  This Pro series advantage can make a noticeable difference in the power needed to run HQ Player effectively.  

If you look at the above table, you can see the strong performance of the RTX A4000 graphics card.  For playing games, the RTX A4000 is considered to be the rough equivalent to an RTX 3070.  Reviews confirm this performance bracket for gaming.  However, for HQ Player, the RTX A4000 slots in closer to the RTX 3090/RTX 3090ti.  Total memory available is also important if you plan on upsampling to DSD 512 with demanding filters such as sinc-L.

 

I hope this information proves useful to others in making their choice of GPU.  Finally, as Miska points out, CPU power is needed for the modulator, so your CPU and GPU should reflect a balanced choice depending upon your goals.

 

 

@Cornell77 would it be possible to share full build of your system? I would also love to know what cooling solution you are using for your CPU and GPU? Also which motherboard? And how many fans in total + the tower case + how much RAM?

[Cornell77]

1 hour ago, shoubhikbhatti said:

 

@Cornell77 would it be possible to share full build of your system? I would also love to know what cooling solution you are using for your CPU and GPU? Also which motherboard? And how many fans in total + the tower case + how much RAM?

Below is a pic of the inside of the PC.  Case is a ZZAW C6. The processor is an I7-11700K, CPU cooler is Noctua-U9S with push pull Noctua fans.  Motherboard is an Asus Z590M. The video card is Nvidea RTX-A4000 (which replaces the ASUS Hyper M2 PCIE silver card (that was previously installed).  The remaining cards are an EVGA Nu Audio card capable of DSD 256 output, Matrix Element H USB card, and Asus Thunderbolt 4 card.  The power supply is a Corsair SFX450.  128GB of Crucial RAM is installed. An Optane 480GB SSD contains the OS. Two Samsung 980 Pro 2TB NVME cards are used for music storage.  There are two top-mounted Noctua 120mm exhaust fans. Each fan has a different blade pitch to better distribute the noise profile. All fans are run on DC and not PWM.

 

I post this here in response to your inquiry.  I hope it helps.  That said, we should ideally return to Nvidea GPU discussion.

[AudioFanatics]

@Rovo Thanks for the information.

@Cornell77 Thanks for the advice. I have a AMD 5950x now, as I intend to create the HQPlayer PC for my 8 channel system on dual purpose (multichannel audio & 4-way digital XO stereo audio), so 8 P-Core of 13700k is possibly not enough.

         I am prepare to get one A4000. The price for used one is pretty good now, and it has more vram than 3080Ti, which is essential for my future system.

 

[NellyWhads]

If anyone in this thread could help me with a quick test I would appreciate it:

I am thinking about getting a GPU to handle up-sampling using HQP, but my goal is to have a completely passively cooled system. The case (HDPlex H5) can handle 95W TDP per heatsink, and has two heatsinks. I can use one for the CPU (to be decided) and one for the GPU. I will be getting the Holo May soon, so the higher the sample rate the better (PCM 1.5M and DSD 512 would be golden. DSD 1024).

 

I do not think I can count on a CPU-only build due to the limited heat dissipation of the cases I want to use (95W TDP). Can push it to 125W TDP, but I would really rather not.

 

Could someone try TDP limiting their A4000 and CPU to 90W and see what is the upper limit of up-sampling possible using one of the intense filters (sinc-L, ext3, xla)? @Cornell77 and @Rovo if you see this and could maybe help with an experiment in your systems, I would truly appreciate it.