NTUplace4dr: A Detailed-Routing-Driven Placer for Mixed-Size Circuit Designs With Technology and Region Constraints

Routability-aware pin access optimization for monolithic 3D designs

“…These two experiments were performed on a 2.4GHz machine with 128GB memory in the C++ programming language based on the 2015 ISPD contest benchmarks [3]. As the inputs of our algorithm are legalized detailed placement solutions, we generated these inputs by NTUplace4dr, the 1 place placer at the 2015 ISPD Placement Contest [7].…”

DREAMPlace 3.0

“…To show the effectiveness of our multi-electrostatics based VLSI placement algorithm, we evaluate the HPWL and congestion metrics on ISPD 2015 contest benchmark suite [1] and compare the results with Eh?Placer [4] and NTUplace4dr [3] using the same target utilization limit. Since Eh?Placer and NTUplace4dr binary executable files currently do not support ISPD 2019 [27] or ICCAD Table 1: Comparison of HPWL, top5 overflow, and runtime with Eh?Placer [4] and NTUplace4dr [3] on ISPD 2015 benchmark suite with region constraints. OVFL is short for overflow.…”

“…It is worth noting that in this work, our placement engine is wirelength-driven without explicit routability optimization. Therefore, only comparing the HPWL metric with Eh?Placer [4] and NTUplace4dr [3], which involve dedicated optimization effects on detailed-routability and design rule violation, etc., is not a very fair evaluation. We choose to show the top5 overflow, the average global routing overflow in the top 5% congested routing grids, reported by NCTUgr [32] global router as a representative metric for routability evaluation.…”

“…A low target density represents a less dense placement solution, and different regions are likely to have different cell distributions, thus different utilization/density limits. Previous placement algorithms [3][4][5] adopt clustering, look-ahead rough legalization methods, and routability optimization techniques to address fence region and density constraints. In the current electrostatics-based placement framework [22], electric forces drive all movable cells to spread towards lower potential energy without considering region constraints or region-specific density limits.…”

“…An analytical global placer models the problem as a constrained nonlinear optimization which determines the cell locations while trying to minimize the wirelength and cell overlap. Previously, NTUplace4dr [3] was proposed to handle region constraints in placement tasks through a region-aware clustering and a new wirelength model. Eh?Placer [4] and RippleDR [5] were proposed to solve region-constrained placement by using an upper-bound-lower-bound optimization method with look-ahead rough legalization honoring region constraints.…”