different time run with gpu

[SabraHashemi]

i runned this code on gpu cuda10.1 pytorch16 when i run first second third its give me 0.22s for each image process(average) but if i run for forth or fifth (i mean sometimes) it gives 0.11 in average for each image why its different ?

[mohit-217]

There is concept of GPU Warmup. So for first request it took time but when GPU starts functioning so it performs well. You can also understand it terms of Overhead cost and first time resource allocation.