Nexus of Sampling, Sequential Decision-making, L2O and Cloud

Today’s IC design optimizations were developed for pre-cloud era compute infrastructure of the 1980s and 1990s. We seek to reinvent CAD optimizations in a {sampling, L2O, federated, cloud} context. We apply distributed sampling and sequential decision-making methods to core optimizations, starting from detailed routing and concurrent placement and routing.

To improve Quality of Results (QoR) outcomes for complex tools and flows, blackbox hyperparameter tuning, or autotuning, has been actively studied in recent years. Autotuning studies in the RTL-to-GDS flow domain have leveraged a variety of search algorithms. However, lack of unified naming and format for metrics incorporated into reward functions limits the potential to apply a single framework across multiple EDA tools. The fact that many frameworks are not open, or have search algorithms strongly tied to the framework, hinders progress by tool users and academic researchers.

An open-source flow parameter autotuning framework for the RTL-to-GDS domain is built on earlier work (Jung et al., 2021). We proposed Early Stopper, a “doomed” detailed routing run prediction v1.0 feature, to enhance the autotuning framework. Early Stopper predicts runtime based on total DRVs per iteration of detailed routing.

Current work includes polishing and production-strength deployment of autotuning, focusing on the DARPA-funded OpenROAD open-source RTL-to-GDS EDA tool, and the Intel Quartus FPGA synthesis, placement and routing tool. We continue to use the Ray/Tune platform.