DALI-PD: Diffusion-based Synthetic Layout Heatmap Generation for ML in Physical Design
This addresses data scarcity for researchers in physical design, though it is incremental as it builds on existing diffusion models for a specific domain.
The paper tackles the limited availability of high-quality training datasets for machine learning in physical design by introducing DALI-PD, a diffusion-based framework that generates synthetic layout heatmaps, creating over 20,000 configurations that improve ML accuracy on tasks like IR drop prediction.
Machine learning (ML) has demonstrated significant promise in various physical design (PD) tasks. However, model generalizability remains limited by the availability of high-quality, large-scale training datasets. Creating such datasets is often computationally expensive and constrained by IP. While very few public datasets are available, they are typically static, slow to generate, and require frequent updates. To address these limitations, we present DALI-PD, a scalable framework for generating synthetic layout heatmaps to accelerate ML in PD research. DALI-PD uses a diffusion model to generate diverse layout heatmaps via fast inference in seconds. The heatmaps include power, IR drop, congestion, macro placement, and cell density maps. Using DALI-PD, we created a dataset comprising over 20,000 layout configurations with varying macro counts and placements. These heatmaps closely resemble real layouts and improve ML accuracy on downstream ML tasks such as IR drop or congestion prediction.