Feedforward Neural Network Models for FPGA Routing Channel Width Estimation

Since interconnects play the increasingly important role in delay and area of the Field-programmable gate array (FPGA) implementations, routing architecture design has become the focus of much work related to FPGA architecture development. This paper leverages feedforward neural networks to derive accurate models of the routing channel width in homogeneous FPGA architecture with two advanced intelligence learning techniques: Gradient-based learning algorithm (GLA) and Extreme learning machine (ELM). The resultant models can be used in the early stages of FPGA architecture development to facilitate fast design space exploration which is difficult to achieve in the traditional experiment-based method. The proposed models are evaluated by comparing the estimated channel widths to the real values generated from a CAD tool VTR over IWLS2005 benchmark circuits. Results show that the GLA model achieves the estimation accuracy 3.98% and the ELM model has the accuracy 3.91%, which show significant improvement over existing estimation approaches.