Search-based synthesis of approximate circuits implemented into FPGAs

Abstract

Approximate computing is capable of exploiting the error resilience of various applications with the aim of improving their parameters such as performance, energy consumption and area on a chip. In this paper, a new systematic approach for the approximation and optimization of circuits intended for LUT-based field programmable gate arrays (FPGAs) is proposed. In order to deliver a good trade-off between the quality of processing and implementation cost, the method employs a genetic programming-based optimization engine. The circuits are internally represented and optimized at the gate level. The resulting LUT-based netlists are obtained using a commercial FPGA tool. In the experimental part, four commonly available commercial FPGA design tools (Xilinx ISE, Xilinx Vivado, Precision, and Quartus) and state-of-the-art academia circuit synthesis and optimization tool ABC are compared. The quality of approximated circuits is evaluated using relaxed equivalence checking by means of Binary decision diagrams. An important conclusion is that the improvements (i.e. area reductions) at the gate level are preserved by the FPGA design tools and thus the number of LUTs is also adequately reduced. It was shown that the current state-of-the-art synthesis tools provide (for some instances) the results that are far from an optimum. For example, a 40% reduction (68 LUTs) was achieved for ‘clmb’ benchmark circuit (Bus Interface) without introducing any error. Additional 43% reduction can be obtained by introducing only a 0.1% error.