you can saturate memory bandwidth without SIMD, since you can issue at least 2 8-byte scalar loads per cycle. it does not leave much room for actual processing though