Search

This paper presents a genetic algorithm hybridized with a constructive procedure and reports its application on the crosstalk aware track assignment problem. In this algorithm, only dominating elements are encoded as chromosomes, on which genetic operators work to explore the solution space, while other elements are determined using constructive method. With proper dominating elements identification, the proposed approach essentially searches a much smaller space without trivial operations, efficiently generating competitive solutions with an effective constructive procedure. Experimental results on a set of industrial instances and ISPD98 benchmarks show that the proposed algorithm reduces both capacitive and inductive coupling in acceptable running time. It is probable that the proposed approach provides a practical way for the application of genetic algorithm on large scale engineering problems.

This paper presents a new floorplanning algorithm emphasizing power reduction for SOC designs using voltage islands. In this algorithm, the supply voltages and positions of blocks are determined simultaneously for both dynamic and static power reduction. Special notice is taken of the interdependence between power and temperature, and thus a block level power and thermal analyzer is incorporated for thermal aware power estimation. Other goals, including area, wire length, as well as level converters and temperature distribution are taken into account, leading to a multi-objective optimization problem, solved using simulated annealing. Experimental results on a set of modified MCNC benchmarks show that introducing voltage islands can reduce the total power by 15% to 30%, and thermal aware voltage island optimization can further reduce the total power by 4% to 15%, as well as promoting even temperature distribution.

In this paper, we present an efficient method to simultaneously size wire widths and decoupling capacitance (decaps) areas for optimizing power/ ground (P/G) networks modeled as RLC linear networks subject to reliability constraints. We formulate the problem as a nonlinear optimization problem and propose an efficient gradient-based non-linear programming method for searching the solution. We apply a time-domain merged adjoint network to efficiently compute the gradients and a novel equivalent circuit modeling technique to speed up the optimization process. The resulting algorithm is very efficient and experimental results show that the new algorithm is capable of optimizing P/G networks modeled as RLC networks with million nodes within 10 hours in modern workstations.