Resumen de Algorithmic techniques for physical design: macro placement and under-the-cell routing
Alexandre Vidal Obiols
With the increase of chip component density and new manufacturability constraints imposed by modern technology nodes, the role of algorithms for electronic design automation is key to the successful implementation of integrated circuits. Two of the critical steps in the physical design flows are macro placement and ensuring all design rules are honored after timing closure.
This thesis proposes contributions to help in these stages, easing time-consuming manual steps and helping physical design engineers to obtain better layouts in reduced turnaround time.
The first contribution is under-the-cell routing, a proposal to systematically connect standard cell components via lateral pins in the lower metal layers. The aim is to reduce congestion in the upper metal layers caused by extra metal and vias, decreasing the number of design rule violations. To allow cells to connect by abutment, a standard cell library is enriched with instances containing lateral pins in a pre-selected sharing track. Algorithms are proposed to maximize the numbers of connections via lateral connection by mapping placed cell instances to layouts with lateral pins, and proposing local placement modifications to increase the opportunities for such connections. Experimental results show a significant decrease in the number of pins, vias, and in number of design rule violations, with negligible impact on wirelength and timing.
The second contribution, done in collaboration with eSilicon (a leading ASIC design company), is the creation of HiDaP, a macro placement tool for modern industrial designs. The proposed approach follows a multilevel scheme to floorplan hierarchical blocks, composed of macros and standard cells. By exploiting RTL information available in the netlist, the dataflow affinity between these blocks is modeled and minimized to find a macro placement with good wirelength and timing properties. The approach is further extended to allow additional engineer input, such as preferred macro locations, and also spectral and force methods to guide the floorplanning search.
Experimental results show that the layouts generated by HiDaP outperforms those obtained by a state-of-the-art EDA physical design software, with similar wirelength and better timing when compared to manually designed tape-out ready macro placements. Layouts obtained by HiDaP have successfully been brought to near timing closure with one to two rounds of small modifications by physical design engineers. HiDaP has been fully integrated in the design flows of the company and its development remains an ongoing effort.
