The landscape of semiconductor design has historically been defined by a significantbarrier to entry, primarily due to the exorbitant costs of proprietary Electronic Design Au-tomation (EDA) software and the restricted access to fabrication-ready Process Design Kits(PDKs). However, a fundamental shift occurred with the initiation of the DARPA-fundedIDEA program in 2018, which aimed to create a fully autonomous, ”no-human-in-loop”(NHIL) RTL-to-GDSII flow capable of completing complex system-on-chip (SoC) layouts inless than 24 hours. 1 This movement toward open-source hardware design has been fur-ther catalyzed by the release of the SkyWater 130nm and GlobalFoundries 180nm PDKs,which allow researchers, startups, and hobbyists to manufacture actual silicon without theburden of non-disclosure agreements (NDAs) or massive licensing fees. 4 Concurrently, theintegration of Artificial Intelligence (AI) and Machine Learning (ML) into these open-sourcetoolchains has emerged as a critical frontier. By leveraging techniques such as ReinforcementLearning (RL) for floorplanning, Graph Neural Networks (GNNs) for timing prediction, andLarge Language Models (LLMs) for hardware description language (HDL) generation, thecommunity is building ”intelligent” EDA tools that not only match but sometimes exceedthe performance of traditional heuristic-based commercial suites.
Chinmay Rozekar (Sun,) studied this question.