RePlAce

The idea

RePlAce takes the ePlace formulation and does the unglamorous work that separates a paper from a tool. Contributions include dynamic step-size and density-penalty scheduling that adapt per iteration instead of following a fixed script; a constraint-oriented local-density function sharpening behavior near macros and dense regions; and, decisively for practice, routability-driven placement: run a fast global-routing congestion estimate periodically, inflate cells in congested tiles so the electrostatic system naturally thins those regions, and iterate until estimated overflow subsides. The result posted state-of-the-art wirelength on standard benchmark suites while producing placements routers could finish, and the code was released openly.

Why you have already used it

This is global_placement in OpenROAD: the flow that grades this site's real-tool problem runs RePlAce's descendant (the gpl module) every time. Its target-density knob is exactly the -density argument discussed in this site's placement-density problem, and its congestion feedback is why lowering density relieves overflow there.

Where you meet it today

In every OpenROAD-flow-scripts run, as the standard open baseline in placement research, and as the immediate ancestor of the GPU (Graphics Processing Unit) generation: DREAMPlace began life by reimplementing exactly this optimization as a neural-network-style computation graph.