EXPERIMENTAL EVALUATION OF A MODIFIED GENETIC ALGORITHM FOR LIBRARY ELEMENT SUBSTITUTION WITH MEASURED METADATA ON AN OPEN FPGA DESIGN FLOW

The efficiency of a modified genetic algorithm (MGA) with threshold selection and two-stage fitness evaluation is studied for designing computer device cores by substituting library elements. A library of 44 parameterised VHDL modules (111 variants in nine categories) was developed; its metadata were measured on the open GHDL, Yosys and nextpnr flow for Lattice ECP5 FPGAs. For the ALU-8 core, all 512 implementation variants were evaluated exhaustively, which provides the true optimum for comparison. With a median of 9 full syntheses, the MGA finds solutions from the top 5.7 % of the design space (Fmax rises from 45.3 to 103.9 MHz), i.e. it needs 4.4–11.8 times fewer syntheses than the classical GA, with a statistically insignificant difference in quality. For the FIR-16 core (|DS| ≈ 5.4·10⁸), the MGA with a median of 21 syntheses outperforms the classical GA with a budget of 40 syntheses in all 25 pairs of runs. Neither algorithm finds the solution that requires simultaneous replacement of eight identical parallel units, and a simple calibration of the fast estimate reduces the area error from 24–89 % to 3–11 %.