Do you collect any statistics on error actually found and/or corrected by ECC memory on your machines? This should be a good indication of whether ECC makes any difference.