An Efficient Design Space Exploration Methodology for Multi-Cluster VLIW Architectures based on Artificial Neural Networks
Mariani G., Palermo G., Zaccaria V., Silvano C.
Article in conference proceedings
Multi-Cluster Very Long Instruction Word (VLIW) architectures are currently designed by using platform-based synthesis techniques. In these approaches, a wide range of platform parameters is tuned to find the best trade-offs in terms of the selected figures of merit (such as energy, delay and area). This optimization phase is called Design Space Exploration (DSE) and it generally consists of a Multi-Objective Optimization (MOO) problem. The design space for a Multi-Cluster architecture is too large to be evaluated comprehensively. So far, several heuristic techniques have been proposed to address the MOO problem, but they are characterized by low efficiency to identify the Pareto front. In this paper, we propose an efficient DSE methodology leveraging neural networks. In particular, an initial design-of-experiments phase is used for generating a coarse view of the target design space; neural networks are then trained and used to refine the exploration, by identifying efficiently the Pareto points of the design space. This process is iteratively repeated until the target criterion (convergence of the Pareto coverage) is satisfied. A set of experimental results are reported to trade-off accuracy and efficiency of the proposed techniques with actual workloads.
Proc. IFIP International Conference on Very Large Scale Integration VLSI - SoC 2008
Rhodes Island, Greece
design space exploration, multi-objective optimization, neural networks, response surface, system-on-chip (SoC), very long instruction words (VLIW)