Performance impacts of superscalar microarchitecture on SOM execution

Abstract

Neural network simulations are notorious for being very time and resource consuming. However, although general purpose microprocessors have improved the performance of these simulations, little is known about which microarchitecture features contribute the most to this performance improvement. In this context, the paper analyzes the performance impact of various microarchitectural mechanisms found in current superscalar microprocessors on the execution of a famous neural network, the SOM algorithm. The conclusion is that the SOM algorithm does not fully benefit from the sophisticated hardware support existing in a state of the art superscalar machine. It is especially true of the memory hierarchy as well as the branch prediction mechanisms.