LMgr: A low-M emory global router with dynamic topology update and bending-aware optimum path search

Abstract

Global routing remains a fundamental physical design problem. We observe that large circuits cause high memory cost1, and modern routers could not optimize the routing path of each two-pin subnet. In this paper, (1) we develop a dynamic topology update technique to improve routing quality (2) we improve the memory efficiency with negligible performance overhead (3) we prove the non-optimality of traditional maze routing algorithm (4) we develop a novel routing algorithm and prove that it is optimum (5) we design a new global router, LMgr, which integrates all the above techniques. The experimental results on the ISPD 2008 benchmark suite show that LMgr could outperform NTHU2.0, NTUgr, FastRoute3.0 and FGR1.1 on solution quality in 13 out of 16 benchmarks and peak memory cost in 15 out of 16 benchmarks, the average memory reduction over all the benchmarks is up to 77%.