The use of single and multiple seed architectures with a natural based micro-architecture exploration algorithm

Abstract

choice of starting point (or seed) is shown to have a significant effect on the delivery speed and quality of the results. The use of the initial conditions of a modified natural algorithm is investigated for providing Delay/Area Trade-Offs for micro-architecture exploration purposes. The Jiggle Algorithm is based on Darwinian and accelerated evolution ideas, and instead of starting with a single seed, a re-generative multiple-seed approach is taken. A set of non-complicated intuitively-derived heuristics are presented that demonstrate that the choice of seed has a substantial effect on the quality and delivery speed of the results. 2: DSE using natural algorithms Natural algorithms, which model natural processes, include among others, genetic algorithms[2], simulated annealing[2][3], and Darwinian techniques[4]. The Jiggle Algorithm described below is based on Darwin's theories of survival of the fittest, which implies (for Ae) never accepting an architecture that is slower and larger than one already found, even if it leads down a path that finds a near-optimum architecture faster than it would have otherwise. In the Jiggle Algorithm for MAe, the Darwinian technique is adapted with a simple heuristic. A different set of heuristics is used to provide various seeds for the natural process. The main characteristic of natural algorithms is that they are experimental in nature, ie. they experiment in the same way nature does in trying to overcome difficulties.