This is a problem with floating point operations happening at a lower level than the error correction you're imagining. In principle, that's not at all necessary. Are you arguing that it's infeasibly expensive to design a chip with operations that are error correctable?