I have just been migrating from CUDA 4 to 5 and installed the just released CUDA 5.0 on Ubuntu 12.04, but performance has gone down drastically. A particular CUDA kernel that used to take 1 second now is taking 6 seconds. It seems that the new driver is the culprit since when I re-installed the old 4 driver performance is regained, but I don’t know why. Has anybody experienced similar problems?
Share
Apologies for taking so long to get back..
I found out that the main culprit was __threadfence(). Not well understanding its use I was writing something like
On the new drivers this was taking a lot of extra time
I am working on GTX680
Thanks a lot
Daniel