I am an REU student at Northeastern University for the summer. We have successfully implemented a scientific computation application on a GPU but now we are trying to implement the same program using two GPU’s rather than one. I don’t know where to start and I was hoping someone could provide some assistance.
The SDK has an example called Multi-GPU, it should get you started. Basically, you need to use two CPU threads and assign each one to a different CUDA context, so each thread uses each GPU.
Is that done in the kernel??
Kernel 's execution environment is GPU and therefore it is unable to handle CPU-side multithreading. you should create such threads in Host-side.
Hope you got it working. Contact me, I may be able to help, I’ve gotten through some multi-gpu issues recently.