The difference is in the operations. With a b tree the software operates on 1 row at a time, but when represented as a series of bit, then in a 32 bit CPU you can process 32 rows in 1 instruction. Which leads to bitmap index over GPU...
Yes, the only problem I found when writing my own bitmap index is they become cumbersome when asking the question: what value has this column for this row. RLE does not help there.
But they are extremely fast when asking the question: what rows has this value on this column. And that is the whole concept, the ability to flip the question.
