Parallel and high-performance computing

A parallel floating-point sum produces a slightly different result from a sequential sum, even though every input is included once. Which explanation is most likely?