Hi Everybody,
I am trying to compute data transfert bandwith between the host and the device (P4 and GeForce GTX 260) using times which are given with visual profiler. My surprise is that the result that I obtained with Visual Profiler is about 1000M/s higher than that obtained by running the .exe oclBandwithTest in the SDK. This is only for the transfert from Host to Device.
The size in bytes is 218080 pixels of unsigned char for a bandwith evaluated at 2452,02 Mb/s, whereas the data size in the .exe is 32Mb for a bandwith of roughly 1139 Mb/s. Considering that the given time by the profiler is 84us, I just made the calculus. I thought the results would be the same.
So I tried to vary the amount of data transfert in my application and the bandwith is not the same for one amount to another one.
My question is : does everyone know how the GPU Time is computed?
Could you explain me the role of the PCIE bus in bandwith computation?
Best Regards,
Grégory
OS : XP, with P4 and GeForce GTX 260