Adaptive threshold non-pareto elimination: re-thinking machine learning for system level design space exploration on FPGAs

One major bottleneck of the system level OpenCL-to-FPGA design tools is their extremely time consuming synthesis process (including place and route). The design space for a typical OpenCL application contains thousands of possible designs even when considering a small number of design space parameters. It costs months of compute time to synthesize all these possible designs into end-to-end FPGA implementations. Thus, the brute force design space exploration (DSE) is impractical for these design tools.

Machine learning is one solution that identifies the valuable Pareto designs by sampling only a small portion of the entire design space. However, most of the existing machine learning frameworks focus on improving the design objective regression accuracy, which is not necessarily suitable for the FPGA DSE task. To address this issue, we propose a novel strategy - Adaptive Threshold Non-Pareto Elimination (ATNE). Instead of focusing on regression accuracy improvement, ATNE focuses on understanding and estimating the inaccuracy. ATNE provides a Pareto identification threshold that adapts to the estimated inaccuracy of the regressor. This adaptive threshold results in a more efficient DSE. For the same prediction quality, ATNE reduces the synthesis complexity by 1.6-2.89x (hundreds of synthesis hours) against the other state of the art frameworks for FPGA DSE. In addition, ATNE is capable of identifying the Pareto designs for certain difficult design spaces which the other existing frameworks are incapable of exploring effectively.