Not necessarily. The x87 implementation (which is relevant to everything PC, prior to SSE) of IEEE 754 internally uses an 80-bit extended precision representation for floating point numbers. Numbers are rounded to 64 bits for loads and stores of course...
Now, what do you suppose happens during a context switch, when a second process wants to use floating point math?
You can save/load 80-bit values to memory with the regular x87 instructions, it's just called "extended precision". And the accelerated "save FPU state" instructions save the full 80-bit state. Otherwise it would be pretty nutty.
Where you can get unpredictable (though still repeatable) behaviour is when your compiler spills into 64-bit memory slots, and does it a little differently with each little modification to the code.
Now, what do you suppose happens during a context switch, when a second process wants to use floating point math?