Reproduce “FlowTune: End-to-end Automatic Logic Optimization Exploration via Domain-specific Multi-armed Bandit” on OpenResearch

About this paper

Walter Lau Neto, Yingjie Li, Pierre-Emmanuel Gaillardon, Cunxi Yu

FlowTune introduces a domain-specific, multi-stage multi-armed bandit framework to automate end-to-end logic optimization for integrated circuits, demonstrating improvements in area, delay, and wirelength for both ASIC and FPGA designs. The framework integrates with industry-standard EDA tools, providing consistent performance across various logic representations.

Key insights

• The initial transformations within a logic synthesis flow exert the most influence on overall optimization outcomes, shaping the effectiveness of subsequent operations.
• Existing generic machine learning methods for EDA often treat synthesis algorithms as black boxes, limiting their ability to leverage domain-specific knowledge for efficient optimization.
• An iterative, multi-stage approach to multi-armed bandit exploration, biased by prior stage results, can significantly accelerate the discovery of optimal or near-optimal logic synthesis flows.

Summary by alphaXiv.