Early-Stage Fast Routability Prediction
I proposed the first systematic study on CNN-based routability prediction. It attracts wide attention in ML for EDA and becomes a common baseline. Some feature-extraction and model-design principles first proposed by us are still widely adopted. Published in ICCAD’18, DATE’19, and also presented in Nvidia GTC’18 and TAU’19. I collaborate with Cadence to validate routability estimation during my internship. Our follow-up works are published in ICCAD’21, DAC’22, addressing the ML model design and data availability challenges in routability prediction. I also write a book chapter on this topic.
Timing and Noise Prediction
I proposed the first GNN-based method for pre-placement net length and timing estimation. It improves the pre-placement slack estimations from commercial tools. It is published in ASP-DAC’21 and TCAD’22. For voltage noise that affects timing and signal integrity, I developed the first design-independent estimator for both vector-based and vectorless IR drop in Nvidia. It reduces IR hotspots by 30% after being integrated into Nvidia’s in-house IR mitigation flow. It is published in ASP-DAC’20 and ICCAD’20 (invited). My collaborative work on crosstalk prediction is also published in ICCAD’20.
Extremely Efficient On-Chip Power Modeling
I developed the first runtime on-chip power monitor that simultaneously achieves per-cycle resolution and less than 1% area overhead without compromising accuracy. The overall framework is fully automated. It is verified on industry-standard CPU cores Arm Neoverse N1 and Cortex-A77. It attracts wide attention within Arm, resulting in three internal presentations and two patent applications. It receives the MICRO’21 Best Paper Award. It is covered in Duke University News, AI Tech Review, etc. My follow-up work for an even more efficient power model is published in ICCAD’22.
Security and Reliability in ML for EDA
I proposed a systematic study on all security and model reliability concerns in existing ML for EDA solutions. I publish the first survey paper on this topic in TCAD’22. My collaborative works about robust ML models and model privacy in EDA applications are published in ICCAD’22 and ASPDAC’23 (Best Paper Candidate). Two more follow-up works were submitted to DAC’23.