A Simple, Fast, and GPU-friendly Steiner-Tree Heuristic

Rectilinear—Steiner-Minimum-Tree (RSMT) generation is a key step in VLSI routing. This paper describes a deterministic RSMT heuristic named SFP because it is simple, fast, and parallel. It is based on a rapid method for computing the optimal solution for three-pin nets, which it extends and applies to larger nets. Compared to FLUTE, a fast and high-quality RSMT algorithm, SFP incurs a wirelength penalty of 0.38 percent on the 16 ISPD 2008 inputs with 13 million nets. However, on a 16-core 2950X CPU, parallel SFP is 19 times faster than FLUTE, and on a Titan V GPU, it is 86 times faster. Serial SFP can generate an RSMT for a 10,000-pin net in under 1.5 seconds.