Comment by monocasa

In practice, branch target predictors are also used for conditional branches too, and even unconditional, direct jumps. Yes, the target can (sort of by definition) be computed from the instruction and the pc, but the by using the branch target buffer, you avoid pipeline bubbles in the correctly predicted case by only having a data dependency on the pc, and predicting the target while the instruction is still being fetched. That way on the next cycle you're already fetching from the predicted target while still decoding the branch instruction itself in the next stage(s).