FusionCell: Cross-Attentive Fusion of Layout Geometry and Netlist Topology for Standard-Cell Performance Prediction
The article introduces FusionCell, a new dual-modality predictor for standard-cell performance prediction. It integrates layout geometry and netlist topology to enhance accuracy and speed in performance characterization. Experimental results show significant improvements in prediction error and characterization time compared to traditional methods.
- ▪FusionCell treats routed layout geometry and netlist topology as inputs and fuses them in a unified model.
- ▪The model uses a DeiT encoder for layout processing and a graph transformer for netlist modeling.
- ▪FusionCell achieved an average MAPE of 0.92 percent and accelerated the characterization process significantly.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.20287 (cs) [Submitted on 19 May 2026] Title:FusionCell: Cross-Attentive Fusion of Layout Geometry and Netlist Topology for Standard-Cell Performance Prediction Authors:Haoyi Zhang, Kairong Guo, Bojie Zhang, Yibo Lin, Runsheng Wang View a PDF of the paper titled FusionCell: Cross-Attentive Fusion of Layout Geometry and Netlist Topology for Standard-Cell Performance Prediction, by Haoyi Zhang and 4 other authors View PDF HTML (experimental) Abstract:Standard cells form the building blocks of digital circuits, so their delay and power critically influence chip-level performance; yet characterization still relies on slow simulation sweeps, and many fast predictors ignore layout geometry, missing coupling and layout-dependent effects.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
