Efficiently Exploring FPGA Design Space Based on Semi-Supervised Learning

摘要

Design space exploration (DSE) is an important step before the physical level design of Field programmable gate arrays (FPGA). An optimum architecture is usually selected from the whole space. As the architecture parameters increase, the huge time cost to explore an exponentially increasing space makes this method unrealistic. We propose a novel predictive modeling approach called ECOMT to estimate the area and delay of a circuit which is mapped onto an FPGA with certain architecture. Semi-supervised model tree is adopted to model the performance with respect to architecture parameters. Combined with nonlinear programming, the area and delay model obtained can be used to guide the DSE. Experimental results show that the model trained through ECOMT has Mean relative error (MRE) below 5% compared to VTR. Meanwhile the time used to attain the model is less than 3 minutes, which reduces the time of DSE considerably.