RALF: A Reinforcement Learning Assisted Automated Analog Layout Design Flow

Jakob Ratschenberger and Harald Pretl

Abstract

The implementation details of an electronic design automation tool using a novel reinforcement learning-assisted automated analog layout design flow (RALF) are presented. The tool generates the layout of an analog circuit with a SPICE netlist as the input, using the SkyWater Technologies SKY130 process design kit and the open-source VLSI layout tool Magic. Two different device placement strategies are implemented, one based on reinforcement learning, and the other based on simulated annealing. Automatic recognition and optimized treatment of structures like differential pairs, current mirror loads, and resistor strings are implemented using subgraph isomorphism. This tool also naturally supports hierarchical circuit topologies. After the placement, a two-stage routing algorithm connects the devices. The first stage is a wire planner, which plans the routing on a rough tile-based grid and uses a negotiation-based algorithm to eliminate estimated routing congestions. The second stage is a detailed router, which facilitates the previously planned guiding to lay out the resources in the routing space. A gridless approach, based on the expansion of obstacles in the routing area, enables this tool to route with variable wire widths as required by user-defined constraints.