I would like to use the CUDA acceleration of my GTX 295. The software that I use, requires sys-cluster/openmpi, which when enabling CUDA requires nvidia-cuda-toolkit, which requires x11-drivers/nvidia-drivers of the same recency, and ditto for GCC. But my GTX 295 is only supported by drivers up to <=nvidia-drivers-340, which is only compatible with older <nvidia-cuda-toolkit-6.5.19, which is not compatible with GCC 4.9+. What is required, is an openmpi that can use the acceleration of my GTX 295. And either enabling USE="cuda" should pull in packages compatible with the GPU, or there should be another way to use the GPU. (Is there a nouveau equivalent? If yes, it would be cool if portage would just pick that then, given the age of the nVidia solution. Otherwise, depending on the GCC 4.8 slot would be an idea too.)
Enabling the cuda useflag in the openmpi package just enables CUDA-aware MPI. There is no automatic acceleration! With a GTX 295 you won't be able to use CUDA-aware MPI and furthermore, the MPI code needs to make explicit use of CUDA. CUDA-aware MPI requires a GPU with compute capability 2.0 or later, thus a GPU of the Fermi generation or newer [1]. [1] https://devblogs.nvidia.com/parallelforall/introduction-cuda-aware-mpi/
(In reply to Marius Brehler from comment #1) Aah, alright. MPI… whatever it is… I don’t care about that anyway. :) Thanks. This information was impossible to know or find, apparently.
