It could be interesting if NVIDIA could add another test mode for the CUDA (programming) Sample (example) called “Bandwidth Test”.
I checked the sample from cuda toolkit 5.5 (utilities).
This sample only tests “hosts to device”, “device to host” or “device to device”. (And variations of pinned/unpinned memory etc).
But it does not actually test the bandwidth from GPU RAM to GPU Processor (reading) or GPU Processor to GPU RAM (writing).
A new mode to test such a situation would make the bandwidth test more complete ?!
Something which will also interest many ! ;)