TimberWolf

The idea

Apply simulated annealing (see the Floorplanning catalog's engine page) directly to cell placement. TimberWolf's move set: displace a random cell to a random location within a range window; swap two cells; flip a cell's orientation. Cost: estimated wirelength (half-perimeter per net) plus penalties for row-length imbalance and overlap, all evaluated incrementally per move. The critical part is the schedule: high initial temperature, geometric cooling, and, TimberWolf's signature, the range window shrinking with temperature, so early moves reorganize the chip globally while late moves polish locally. Later versions refined the schedule empirically and added row-based bookkeeping that kept per-move evaluation cheap.

Its reign and its ceiling

Through the late 1980s and the 1990s TimberWolf and its descendants were the quality standard; when runtime was affordable, nothing beat a long anneal, and version after version won benchmark comparisons. The ceiling was arithmetic: acceptable schedules need move counts that grow super-linearly with cell count, and at hundreds of thousands of cells the wall-clock became indefensible against analytical methods that solve for all positions simultaneously. Annealing retreated from general placement to the niches where its indifference to cost-function shape still wins.

Where you meet it today

Detailed-placement refinement on small windows, macro packing and analog placement (where cost functions are lumpy and object counts small), FPGA (Field-Programmable Gate Array) placement in the classic VPR (Versatile Place and Route) tool, and every course's placement chapter as the canonical annealing case study.