Accelerating Scientific Model Optimization with a Pipelined FPGA-Based Differential Evolution Engine

Abstract

Differential Evolution (DE) [5] with Numerical Integration (NI) is an ideal target for Custom Computing Machines on FPGAs, since it produces deep pipelines, requires minimal external memory bandwidth, and benefits from large memory bandwidth. DE is a genetic algorithm used for scientific model optimization. We propose a generic FPGA-based DE architecture, parameterized to accommodate to different scientific models. It supports both non-adaptive and adaptive NI methods. The core DE engine is programmed in VHDL for high adaptability and performance, whereas the scientific models and their NI are programmed in C++ for flexibility and easiness of development.