A regular branch predictor try to guess which way a branch (eg: if-else) will go before it's executed. That way, CPU can fetch and decode instructions in advanc

A regular branch predictor try to guess which way a branch (eg: if-else) will go before it's executed. That way, CPU can fetch and decode instructions in advance. Each side of the branch leads to the start of a new block of instructions. The last instruction of such a block is usually another branch. In other words, a branch predictor guess the address of the next block. A 2-ahead branch predictor does the same thing, but for the two subsequent blocks. As stated in the paper: "information from the current instruction block is used for predicting the address of the block following the next instruction block". Unlike a regular branch predictor, it can do that without having to wait for instructions in the next block to be decoded. That way, you can feed multiple instruction decoders at once. This is particularly useful in modern CPUs where the instruction decoder has become a bottleneck. With only 1 decoder and 1 instruction decoded per cycle (at best) it cannot cope with a wide frontend that can execute many instructions (eg: 4-6) per cycle.