I have a Tesla C870 card installed on Fedora Core 7 x86_64 using the driver NVIDIA-Linux-x86_64-169.07-pkg2.run. When I run the CUDA SDK bandwidthTest i get the following numbers:
Host to Device 975.4
Device to Host 798.8
Device to Device 63923.3
Does this seem correct? Am I using the right version of the driver? I saw in other posts on these forums that the 169.09 version had issues causing slow memory, which is why I have installed the .07 version. Hopefully I will soon have time to plug our S870 in to see how it looks too. ;)
On a somewhat unrelated note, do the CUBLAS functions block until the result is computed on the board? In other words, if I call gettimeofday() immediately before and after a call to cublasSgbmv(), will the difference give me the execution time of that function?
Thanks for any help!!