On mainframes where it really, really matters that the answers be right, three processors will do the same calculation in lockstep and any one processor deviating will have its results thrown out.
If I read the article right, it's not that it sometimes fails, rather it almost always fails, but the amount of error is predictable. You can converge on the exact value by repeating and averaging the calculation, but it would be simpler to have an exact co-processor.
