Why the inference speed differs a lot between windows and linux on pytorch-cpu

why the inference speed differs a lot between windows and linux on pytorch-cpu?
linux is faster than windows using the same pytorch cpu version without cuda?
why why why!

really important to me!
hope i can get the answer

There is some inherent windows problem with either L2 cache or memory page management, this difference is not specific to pytorch. Also, your builds may be not identical in regards to MKL, avx, openmp threading.