I got a response to my blog from Sascha Pfaller who's writing another MD5 brute forcer. As it seems there is a faster open source alternative coming up!
After skipping through this new code I realized that I forgot to interlace my ROTATE_LEFT functions. As this is actually 3 operations each, VC compiler failed to interlace this for me. I now went from 117 Mhashes/s to 135 Mhashes/s on my system with VC compiler. Apparently Intel compiler already interlaced my ROTATEs before, no speed increase with my new code with Intel compiler.
I also tried to fix my source code to work with Linux. I still have some stupid problem with pthreads and signaling conditions or something. I give up for now and just release a version that at least works. If you plan on using it, mind the README:
Linux support is in a state of 'it works and should find your plaintexts'. There are some bugs; Speed doesn't always show correctly or sometimes not at all. If you specify more threads then you have cores in your system, using -t, you might see high numbers for the speed. Your system is not actually running that fast, sorry :)
New version of EnTibr, Cacheebr and shabr might be coming up if anyone actually wants to use those in Linux.
Please feel free to leave comments or feedback about how fast you are going with the various versions. Or if you notice any bugs, please let me know!