performance - What's the speed of texture upload? -

i upload 2 images gpu memory, , i'm interested how fast can this?

in fact - faster compare 2 bitmaps in ram cpu, or upload them gpu , use gpu parallelism it?

if run cuda device bandwidth sample, you'll benchmark upload speed.

assuming ddr3 tri-channel 1600mhz ram, you'll 38 gb/s memory bandwidth.

take typical midrange card gtx460 , you'll 84 gb/s memory bandwidth. note you'll have make hop across bus 8gb/s theoretical, ~5.5 in practice pci-e2.0 x16 link.

note kotlinski's answer isn't quite correct. you'll can compared in parallel , parallel reduction in case, bigger gpu device bandwidth can work win out eventually.

i think answer be: loss upload gpu , comparison once. possible gain if comparison made multiple times (kept , modified on gpu, example).

edit:

the multiple times comparison refers if modified images on gpu memory in situ. thus, merit comparison (caching doesn't cut it), while not incurring penalty of copy across bus.