Placement

The paper separates GNN approaches from differentiable placers in placement and floorplanning. Differentiable placers are algebraically matched to placement objectives but are not GNNs. AlphaChip’s public production status is treated as important but qualified by evaluation and benchmarking concerns. Placement is framed as a continuous coordinate optimization problem with hypergraph wirelength and density penalties. AlphaChip uses a graph encoder with reinforcement learning and optimizes a reward based on normalized HPWL, congestion, and density.