ReLS: Retrieval Is Efficient Knowledge Transfer For Logic Synthesis
Year: 2024
Type: conference-paper
Abstract: Design houses manage extensive collections of legacy synthesis data containing valuable knowledge crucial for enhancing synthesis efficiency and performance. Effectively encapsulating and transferring this knowledge is a pressing research challenge. Previous methods incorporate synthesis knowledge into neural networks trained on legacy data, often necessitating retraining for updated databases and fine-tuning for new designs. This study introduces ReLS, a retrieval-powered logic synthesis framework that integrates network weight-captured knowledge with information stored in dynamic synthesis databases. ReLS employs a graph neural network-based And-Inverter Graphs (AIGs) encoder to convert new AIGs into vector representations for similarity searches against legacy AIGs. Top-ranked recipes from similar AIGs are retrieved and re-ranked to offer initial solutions for advanced synthesis optimization algorithms. ReLS serves as an initialization tool to streamline efficient knowledge transfer without the requirement for fine-tuning. Utilizing the OpenABC-D database with 61, 500 synthesis samples and limiting to 50 synthesis runs for each design, ReLS enhances the quality-of-results improvements of various existing methods over handcrafted recipes by 8.6% to 196% on 10 unseen designs. Furthermore, experimental results showcase ReLS's ability to harness databases not used during training.
Institution: Nvidia (United States)
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4402172467