A Fast Parameter Tuning Framework via Transfer Learning and Multi-objective Bayesian Optimization
TimeTuesday, July 12th2:15pm - 2:37pm PDT
Location3007, Level 3
System-on-Chip Design Methodology
DescriptionIn VLSI, it is difficult to manually modify the configurations to achieve the optimal PPA objectives. Design space exploration can solve this problem effectively. However, the prior knowledge of the implemented technologies is neglected. In this paper, a fast parameter tuning framework via cross-technology transfer learning is proposed to quickly find the optimal configurations. Firstly, GCS is utilized to transfer prior knowledge. Secondly, TL-MOBO is used to search for Pareto-optimal frontier efficiently. Finally, UASAF is employed to find Pareto configurations accurately and selects the next candidate. Experimental results show that this framework can achieve better Pareto-optimal frontier than state-of-the-art methodologies.