BRANCH PREDICTION SIMULATION BY USING NEURAL PERCEPTRON IN SIMPLESCALAR

• Main Author: INAWEL (NIM: 132 04 046); Pembimbing : Ir. Yudi Satria Gondokaryono, M.Sc., Ph.D, ILHAM
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/16007
• Institution: Institut Teknologi Bandung

Developments in today&#8223;s technology is very fast, one of which is computer technology. Currently, computers are no longer foreign and expensive tools, nearly all fields require a computer as a tool, because it has the advantages in terms of speed and accuracy. Instructions in a computer microprocessor is handled in two steps, namely decoding and execution. A microprocessor can save time by doing the decoding of an instruction while executing the instructions that preceded it, this method is called pipelining. In this situation the processor can work on the steps of other instructions at the same time, so several instructions can be executed in a short time, and improve processor performance. <br /> <br /> <br /> Now the problem is that if a branch instruction is found. Branch instruction is the implementation of the if-then-else form. This causes delay in the flow of instructions through the pipeline because the processor does not know which instruction must be executed before the branch instruction is completed. The longer the pipeline is, the longer the delay in the flow of instruction in the pipeline will be before the processor knows which instruction should be inserted later into the pipeline. Because the average length of pipeline on modern microprocessors is lengthy, a solution is needed to this problem. <br /> <br /> <br /> In this final task, a simulator-package called sim-bpred from Simplescalar is modified. The goal is to add another branch prediction simulation function called neural perceptron method to be used as an accurate branch predictor. Perceptron assigns weight vector for each object in a system where each weight represents the strength of the relationship between a perceptron on an object with a perceptron on the other object. The stronger the relationship, the greater the weight that represent it. <br /> <br /> <br /> To test the modifications, a simulation is done by using tools from the SPEC 1995 benchmark, namely Dijkstra, Compress 1995, the Secure Hash Algorithm, and Go. From the test results, it can be seen that branch prediction using neural perceptron has a higher accuracy compared with other existing branch predictor in the sim-bpred.