Push the hyperthread count up higher and you have enough stuff to do while waiting on memory that branch prediction doesn't matter as much. Could call it a GPU.