Reinforcement-Learning Floorplanning (AlphaChip)

The idea

Frame macro placement as a sequential game. The state: the netlist (as a graph), the partially placed floorplan, and which macro is next. The action: choose a grid cell for that macro. After all macros are placed, standard cells are spread by a fast force-directed method, and the episode's reward is a mixture of negative half-perimeter wirelength and congestion and density penalties. A policy network, with a graph neural network encoding the netlist and a convolutional view of the occupancy grid, is trained by reinforcement learning across thousands of episodes, and, the headline claim, across designs: pretraining on a corpus of chips lets the policy zero-shot or quickly fine-tune on a new one, getting better with experience the way annealing never does.

The reception

The Nature 2021 paper reported human-competitive or better macro placements in hours and use in Google's TPU (Tensor Processing Unit) production flows; the code was released as circuit training, and the approach was later branded AlphaChip. It also ignited one of EDA's (Electronic Design Automation's) sharpest reproducibility debates: independent evaluations (notably by Cheng, Kahng, et al.) reported that strong classical baselines, tuned annealing among them, match or beat the published results on open benchmarks, and disputes over preprocessing (netlist clustering, placement seeds) and baseline strength ran through journals for years. Read both sides.

Why it matters regardless

Whatever the final scorecard, the work moved the field: learned placement is now a standing research area, the state-action-reward framing of floorplanning is standard vocabulary, and every major EDA vendor has since shipped some machine-learning-guided floorplanning or design-space exploration. It is the first floorplanning algorithm whose skill is a trained artifact rather than a coded procedure.

Where you meet it today

The open-source circuit training repository, vendor reinforcement-learning offerings, and a lively benchmark literature (MacroPlacement, among others) that this controversy productively forced into existence.