200 Million addresses loaded and working with CPU; GPU untested
It eats up 4300MB, 4.3GB of RAM. Binary to Bloom; I am sure if it was loaded via text file, it would eat up a lot more. Also, I doubt this will work with Bitcrack or any VanitySearch forks.
Very impressive stats you found but I highly doubt you will get any higher than this if you test on GPU, this is because of its architecture where it's memory is separated from all the rest of the system RAM.
So you could have a 96GB system (eg. A very very recent MacBook Pro with an external nvidia GPU (does that even exist??) but the GPU will only have only 8 or 16 gigs total which puts a ceiling on the number of addresses. That's kinda sad as these suckers can easily do 500x the search performance of a single CPU socket.