Indeed, 53 significant bits (doubles) vs 24 significant bits (floats). In my case float performance was what I needed. There are always other options, depending on your (precision) needs and needs in general, such as binary coded decimals, fixed point arithmetic...