Well, yeah. The binary was built for processors with AES-NI - there have been no improvements to non AES-NI code yet.
EDIT: Also, the little to no increase in Windows is likely due to Windows having no mmap() function. Therefore I cannot use hugepages (or prefault aforementioned hugepages.)
If you want to damage your brain (even more than if you were to continue your sleep-depravation routine
) you could try out using CreateFileMapping / VirtualAlloc and friends on Windows to grab some huge pages. The brain will start to hurt when the windows users complain about it not running without them doing some setup work.