How to communicate beetween two GPUs Tesla D870 : two tesla C870 GPUs

Hi everyone !

I’d like to know if there is a way to tranfer data directly from a GPU to another without using an intermediate space on the host’s ram ?
If so, where may I find some informations ?

I think it could be a good feature to simplifyprogrammation of datas transfers . The only way I can find to do this uses space on the host ram to store the data and then sending back the results to the other card. So it needs synchronization mechanism beetwween threads (quite easy on this simple case but still…) but more important : it needs twice the bandwidth to do this…

Do you have any idea ?



This is not possible in CUDA 1.1.


Thanks for the answer. I’ll do it the other way ;)