how is it even possible? You can't hash 2 different block headers and spend only 1/2 of the computing power of hashing only one of these blocks...
It's already working on the testnet, so you don't have to understand anything - just know someone figured out a way of doing it.