Compiling and Executing OpenACC on TESLA and GEFORCE

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hello
Hope you are doing well.
Can you kindly help me understand why my Fortran OpenACC code could not detect a Tesla K40c GPU and a Geforce GTX Titan X at the same time? (They are both on the same desktop/pc - running windows 10)
I understand that the flags to run OpenACC code on both is -ta=tesla:cc35 and -ta=tesla:cc50 respectively. However, I had the Titan X working well with my code and cc50, but when I run with cc35, the K40c is not even visible when I try to search for it with openacc functions.
Is it possible that I cannot run a single code that can make use of both GPUs (or even making use of Multi-GPU) depending on user choice? I would like to let the person who executes to choose between them.
I had a workaround, that is hefty. I downloaded these two drivers by NVIDIA:

  1. 416.16-desktop-win10-64bit-international-whql (for the Titan X)
  2. 398.75-tesla-desktop-win10-64bit-international (for the K40c)
    When I install one of them, I am able to run my OpenACC code with the respective GPU.
    Could you kindly inform me of the problem here?
    Thank you for your time
    Ahmed
2

Hi Ahmad,

This sound more like a CUDA driver issue. Since the Titan X is a gaming card, it will be using the WDDM driver while the K40c would use the TCC driver. Granted, this is out of my area of expertise, but I don’t think you can mix the two types of CUDA device drivers on the same system.

If they were both the same type of cards (Tesla vs GeoForce) or you were using Linux (which only uses TCC), then you can mix the compute capability by using “-ta=tesla:cc35,cc50” when compiling the application.

-Mat

1 Like
3

Hi, I have a HP Z800 workstation. I have a tesla K40c, 2 Tesla K8 for GPU rendering and a Quadro M5000 for graphics and computing.

Works amazing for its age.

Then I switched the M5000 for a GTX Titan X 12gb, since the K40c also has 12gb of ram.

But I found myself on a driver’s loop. It sounds like I am in the same scenario as Ahmed, can you please confirm they are incompatible for cuda gpu rendering? I am running Blender 4.2 on Windows 10 OS

Thank you

4

Hi Waracana,

Sorry, but I haven’t worked on Windows for many years so am not up to date on the driver situation. Also since I support the NVHPC compilers and directive based programming models, the advice above was just from my experience as a user, not as an expert.

I did find this video which may be helpful: https://www.youtube.com/watch?v=5nLhKhnQRjo

Seems that the solution is to use a virtual GPU to enable the data center cards to use the WDDM driver. However, it doesn’t appear that these older K40c cards are supported via this method. He also mentioned a cost associated with the vGPU, so it may not be free. See: Virtual GPU (vGPU) Software Free 90Days Trial | NVIDIA

There is a vGPU forum where they might be able offer ideas for a K40c: Virtual GPU Forums - NVIDIA Developer Forums

-Mat