The reference implementation is Python, so there's lots of room to optimize. I wonder if you can eventually build an ASIC that replaces the disk based lookup with in memory calculations of the same data.