In my project we’re doing some OpenGL rendering, and then the result goes to much longer CUDA computations, possibly on a different GPU. Lastly we run across the problem of Session 0 isolation (using CUDA with Windows Services), which is solved with TCC drivers (for Tesla GPUs of course).
AFAIK, the TCC driver solves the problem of Session 0 isolation, but it doesn’t provide support for OpenGL hardware acceleration. So, my idea was to have a workstation with Tesla GPUs (for CUDA) and non-Tesla GPU (e.g. GTX 480 for OpenGL stuff). Nevertheless, the release notes to Tesla driver 260.61 says: “Nonâ€supported NVIDIA GPUs [those without TCC, if I understand correctly] can still function as CUDA devices, but the GPU’s graphics functionality is not available to applications.”
My question is: what is the aforementioned “GPU’s graphics functionality”? Is it OpenGL hardware acceleration, or only OpenGL interop? Does it affect only kernels (e.g. texture fetches, interpolation, …)?
Could someone clarify this to me, please? We have no access to Tesla-GPU.
Thanks, Jack, for your reply. I am aware of the WGL_NV_gpu_affinity extension and surely it will be useful.
I’ve been thinking a bit more about my problem and I have the following conclusions:
The “Nonâ€supported NVIDIA GPUs [those without TCC, if I understand correctly] can still function as CUDA devices, but the GPU’s graphics functionality is not available to applications.” from the Release Notes describes only the case when I’d like to use GTX480 as a CUDA device. In fact, I want to run my CUDA code on Tesla devices, thus the mentioned limitation does not apply to my project.
In the same Release Notes it says that “TCC mode does not include support for 2D or 3D graphics functionality.” If I understand correctly, my kernels will not work if I use texture fetches in my code, right? Could someone confirm that, please?