That's a claim that the authors aren't making. As has been repeatedly pointed out, the attack depends on timing the execution of blocks of code which take much longer than a single instruction, so the 44.1kHz sampling rate described in the paper is sufficient.