FusionCell Predicts Chip Performance from Layout and Netlist
FusionCell is a dual-modality predictor that jointly processes routed layout geometry and netlist topology for accurate standard-cell performance prediction. It uses a DeiT encoder for three-layer routed layouts and a graph transformer for heterogeneous device/net graphs, integrated via a topology-guided mechanism where the netlist queries relevant physical regions. This approach captures fine-grained spatial details and structural connectivity, addressing limitations of slow simulation sweeps and layout-ignorant fast predictors. The work is published on arXiv as preprint 2605.20287.
Key facts
- FusionCell is a dual-modality predictor for standard-cell performance.
- It processes routed layout geometry and netlist topology as inputs.
- A DeiT encoder handles three-layer routed layouts.
- A graph transformer models heterogeneous device/net graphs.
- Modalities are fused via a topology-guided mechanism.
- The netlist acts as a structural map to query layout regions.
- It aims to replace slow simulation sweeps for delay and power prediction.
- The preprint is arXiv:2605.20287.
Entities
Institutions
- arXiv