If I understand correctly the reason for that claim is they know in advance how many bits they will need to essentially guarantee a good sample, so instead of needing to perform the operation first to find out where the next operation can start they can just chunk it out and do it in parallel.