Isn't the limit of supercomputing more about the data link speed than about the actual node compute speed?