A bloom filter needs to multiple loads from different memory locations for each single lookup. (7, in the example 1.2GB filter.) But unlike, say, with a tree, it knows all the addresses after computing the hashes, without having to wait for results from the previous loads. So it can start all of them in parallel.