nvidia - CUDA - how much slower is transferring over PCI-E? -
if transfer single byte cuda kernel pci-e host (zero-copy memory), how slow compared transferring 200 megabytes?
what know, since know transferring on pci-e slow cuda kernel, is: change if transfer single byte or huge amount of data? or perhaps since memory transfers performed in "bulks", transferring single byte extremely expensive , useless respect transferring 200 mbs?
hope pic explain everything. data generated bandwidthtest in cuda samples. hardware environment pci-e v2.0, tesla m2090 , 2x xeon e5-2609. please note both axises in log scale.
given figure, can see overhead of launching transfer request takes constant time. regression analysis on data gives estimated overhead time of 4.9us h2d, 3.3us d2h , 3.0us d2d.

Comments
Post a Comment