Calder et al. [1]suggest improvements to the BTB architecture to make it simpler, at the same time, not compromising with the efficiency of the BTB. They use the branch execution penalty metric defined by Yeh et al. [12] to compare one architecture to another. The penalty for misfetching is less than the penalty for misprediction. The percentage of misfetched branches (%MfB) and the percentage of mispredicted branches (%MpB) are recorded and the BEP is defined as
which reflects the average penalty suffered by a branch due to misfetch and misprediction. However, this binds a particular value for these penalties. In [1], it has been assumed that misfetch penalty is one and misprediction penalty is five cycles, which may constrains the extrapolation of the inferences to all systems.
Based on their simulation study and BEP evaluation of the previous architectures, Calder et al. suggest two changes to the BTB, that may lead to better performance and easier implementation - decoupling branch prediction from the BTB and allocating the BTB for only taken branches. They get a performance that varies with benchmarks, because of their assumptions, not going well with all programs.