I see there are some people who suspect cheating here (which is sad, given that no one has actually verified the accusations yet). It should be easy to find out for anyone with at least two different video cards. Just do a benchmark, then swap in the other card and see if the results change considerably.
On another note, I've seen bits of the original source code myself. Believe me, most people wouldn't understand a thing. All functions and variables have crazy_long_non_english_naming.
On a third note, I can actually imagine that TitaniumGL is faster than LLVMpipe, given that LLVMpipe probably tries to be reasonably accurate and emulate shaders. TitaniumGL probably makes a hell of a lot of shortcuts (which is a perfectly valid approach, if the unsued functions are not/seldomly used, why not leave them out?). Either way, more testing needed here.