Your use of the word "must" is the key. The digits of pi are supposed to be the same, but if a cosmic ray flip a bit, or if the machine becomes unstable because it runs hot, or if there's a bug in the FPU (such as happened to Intel) then results can differ, even when they shouldn't. Computing pi to very large numbers of digits is a standard warmup to make sure that new hardware is working correctly. It's only one of many tests, but it has found hardware bugs.

I remember using a machine that started to give incorrect answers when it ran too hot, and it ran hot when it was working full speed. We had to tweak the compiler to add extra NOPs into the instruction path becuase the micro-code used less power in a NOP, thus generated less heat.

EDIT: I've realised this is another example of the sofware fixed the hardware: http://news.ycombinator.com/item?id=1031384