I have a Tesla and Quadro running in parallel using SDL_Threads on Linux (CUDA 4.0, Tesla C2060, Quadro FX5800).
The Quadro does nothing but generate a 3d texture that’s 480x384x384 using glTexImage3D while the Tesla is doing a bunch of math.
Running them in parallel is no faster than running them one at a time… that is, the Quadro’s texture generation slows down the stuff running on the Tesla.
I don’t see any great reason for it… the memory transfers don’t seem to be bottlenecking the process.
Any hints? I’ve tried OpenGL+CUDA integration and I’d say that it runs even slower.