Efficient FPGA Routing using Reinforcement Learning

With every new generation, Field Programmable Gate Arrays (FPGAs) are getting more complex and so are their back end flow. Routing is an important step of FPGA back end flow that takes a lot of time. Making it more efficient in terms of execution time without the loss of quality is a huge challenge. In this work, we propose to use Reinforcement Learning (RL) based routing technique to make the FPGA routing faster. We use a comprehensive set of homogeneous and heterogeneous benchmarks to compare the RL-based technique with the conventional negotiated congestion driven routing technique. Experimental results reveal that for quick turn around, when compared to negotiated congestion technique, the RL-based technique gives, on average, 35% more accurate results about the final design. Moreover, for the complete routing step, the RL-based technique gives 30% speed up while giving similar quality of results.