That's interesting, makes me want to see the graph of power vs cost for extra error correcting data but with longer refresh cycles on the memory (for that suspend-to-RAM application).