Best Nvidia CUDA Card for HQPlayer

[yamamoto2002]

This is my result of GTX Titan Black, the same GPU die is used for Titan supercomputer. Double precision performance is 1700 Gflops. I think this is still the best card for home GPU computing

 

This is my result of GTX 1080. Double precision performance is about 300 Gflops.

 

On Nvidia control panel, Double precision function of the Titan GPU should be enabled (Default value is disabled. I don't know why). It seems there is no double precision menu appeared on GTX 1080.

[yamamoto2002]

Dual-GPU cards such as GTX Titan Z, GTX 690, GTX 590 have two separated global memory for each GPU and it is a bit more difficult to program than conventional single GPU cards. It is recommended to check if the GPGPU application you runs on it supports such a exotic cards before you purchase

[yamamoto2002]

It seems double precision arithmetic of Titan V card is 22 times faster than GTX 1080.

[yamamoto2002]

This is my Titan V result.

About 6 TFLOPS doubleprec, One-seventh of Earth Simulator Gen1 Supercomputer .

I hope upcoming Geforce Volta products may have some doubleprec capability

 

[yamamoto2002]

On 6/12/2018 at 11:31 PM, louisxiawei said:

What a beast! Have you run any HQplayer heavy filter setting with it? Something like upsampling 44.1/16 → 48 x 512 using poly-sinc-xtr filter?

 

 

 

 

No.

 

It seems, in order to run CUDA programs on Volta, programs should be compiled using latest version of CUDA Toolkit, which dropped support of older Fermi based GPUs such as Geforce GTX 580 or Quadro 6000. I'm not sure this affects HQP

 

> Meanwhile AMD and Intel are having the "multi-core" CPU competition. All good for HQplayer

 

Yes it is good thing. On Windows, non-processor-group-aware apps can handle up to 64 core (or 64 hyper-thread). Process affinity mask is 64bit (one bit is associated to one core(or hyper-thread), so it can express up to 64 core(or hyper-thread)). With 32 core 64 thread CPU, all the available affinity mask bit is used and free performance improvement of multi thread app by increasing CPU core ends there. If this trend continues and say 64 core 128 thread CPU is arrived, app should be rewritten to use multiple processor groups to squeeze all the CPU resource.

[yamamoto2002]

3 hours ago, Miska said:

But the CUDA-Z test application you are running is compiled against CUDA 5 or 6 or something really old.

 

Thanks for your reply. I understood about CUDA binary forward compatibility and things are cleared up:

 

https://docs.nvidia.com/cuda/volta-compatibility-guide/index.html

Quote

 

  Applications that already include PTX versions of their kernels should work as-is on Volta-based GPUs. Applications that only support specific GPU architectures via cubin files, however, will need to be updated to provide Volta-compatible PTX or cubins.

 

 

So, CUDA-Z contains PTX binary for forward compatibility and if you are lucky enough, PTX code runs well on future architecture.

 

[yamamoto2002]

On 6/15/2018 at 6:29 PM, Miska said:

Luckily I have very little Windows specific code and no limitations for number of CPU cores. So no need to rewrite anything...

 

Last I checked, on Windows, one process can handle up to 64 HT. So, in order to handle 128 HT, another worker process should be created and each process create 64 threads, and two processes communicate with inter-process communication. This is significant rewrite from casual multi threading code of Sunday programmer

[yamamoto2002]

Thank you for your reply. I found Linux has much better multi threading support for casual Sunday programmer. Also there are several cross platform library to do overcome this kind of OS specific quirky. And sorry for topic drift.