Thecontinued evolution of VLSItechnology at and belowthe3nmnodeintroduces unprecedented challenges for modern electronic design automation (EDA), especially
in bridging the gaps between timing prediction, physical design flexibility during ECO routing...
Thecontinued evolution of VLSItechnology at and belowthe3nmnodeintroduces unprecedented challenges for modern electronic design automation (EDA), especially
in bridging the gaps between timing prediction, physical design flexibility during ECO routing, and efficient device-technology co-optimization. This dissertation presents an integrated framework that combines deep-learning-based timing optimization, adaptive pin pattern modification for ECO routing, and a new standard cell methodology
leveraging Flip-FET (FFET) technology. Together, these works collectively enable robust, scalable design solutions that address the rising complexity of advanced nodes
while delivering Power, Performance, and Area (PPA) benefits. The first study introduces DTOC-P, a deep-learning-driven timing optimization framework that tightly integrates with commercial EDA tools. DTOC-P enhances the practicality of machine learning-based timing prediction by supporting continual learning within an evolving design environment and employing anomaly detection to filter out outliers. With a hierarchical neural model capturing pre-route timing metrics, DTOC-P guides place-and-route stages in real EDA flows, improving path delay prediction and reducing violation counts. Extensive experiments demonstrate that DTOC-P yields up to 55% error reduction in arc delay prediction and 47% improvement in arc slew prediction, while achieving notable optimization speedup and reliability gains over retraining-from-scratch approaches.
The second study presents an adaptive pin pattern modification (APM) methodology for ECO routing in advanced nodes, addressing the growing complexity of routing congestion and pin accessibility. Unlike conventional ECO router, our APM enabled ECO router dynamically alters standard cell pin patterns through three techniques—PST (pin-shift-trim), PF (pin-free), and PB (pin-bridge)—directly during the routing process. This approach enables the resolution of otherwise unroutable nets, increases DRV fix rate by up to 17.7% over existing tools, and accelerates the ECO process by up to two orders of magnitude, all without significant PPA penalty. The work establishes a new paradigm where cell-level modifications and chip-level routing are co-optimized in real time for maximal design closure effectiveness. Finally, the third study explores a design-technology co-optimization (DTCO) methodology using Flip-FET (FFET) standard cells that fully exploit dual-sided interconnects in three-dimensional device architectures. The proposed FFET standard cells feature reduced cell height, balanced frontside and backside pin accessibility, and optimized internal routing. Through a systematic flow, the DTCO framework minimizes area and wirelength, eliminates tap-cell overhead, and reduces more design rule violations compared to conventional methods. The FFET-based approach not only enables successful scaling of high-density chips at aggressively small nodes, but also demonstrates industrial applicability across multiple real OpenCore benchmarks by integrating with state-of-the-art commercial EDA flows.