EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

GRAPH-AWARE MULTI-TASK LEARNING FOR EARLY PREDICTION OF TIMING VIOLATIONS AND ROUTING CONGESTION IN DIGITAL IC PHYSICAL DESIGN

HaiJian Zhang

World Journal of Information Technology · 01.01.2024 · 2 (1) · 13-20

DOI: 10.61784/wjit4124 ↗

Kurz gefasst

Fachartikel in der Zeitschrift World Journal of Information Technology, online erschienen am 01.01.2024. Verfasst von HaiJian Zhang. Die Arbeit wurde bisher 3-mal zitiert (Crossref). Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Routing congestion and timing closure are the two dominant, and physically coupled, bottlenecks of digital integrated-circuit physical design, yet machine-learning predictors have almost always targeted them in isolation. We present GAMT-Net, a graph-aware multi-task network that jointly predicts per-cell routing congestion and per-cell timing slack from a single netlist-and-placement graph at the pre-routing stage. Its core novelty is a congestion-conditioned timing (CcT) module that feeds the timing head with the predicted congestion embedding of each cell together with the mean congestion embedding of its fan-in-cone predecessors, explicitly encoding the congestion-detour-delay causality that links the two tasks; task losses are balanced by homoscedastic uncertainty weighting. Because large public layout datasets are access-gated, we construct a fully reproducible, physically-grounded benchmark of 220 synthetic designs (166,822 cells) whose labels are produced by a rectangular-uniform-wire-density congestion estimator and a longest-path static-timing engine with congestion-dependent interconnect delay. On held-out designs, multi-task learning improves slack correlation over single-task learning (0.510 vs. 0.482), and the CcT module delivers the largest single gain, raising slack Pearson correlation to 0.543 and violation-classification AUC to 0.806, while congestion prediction reaches r = 0.86 and hotspot AUC 0.89. Ablations show that graph structure is essential, that a naive graph-convolutional encoder over-smooths the directional timing signal, and that netlist connectivity rather than spatial proximity drives timing. All code and data are released for reproducibility.

Zitieren

APA

Zhang, H. (2024). GRAPH-AWARE MULTI-TASK LEARNING FOR EARLY PREDICTION OF TIMING VIOLATIONS AND ROUTING CONGESTION IN DIGITAL IC PHYSICAL DESIGN. World Journal of Information Technology, 2(1), 13-20. https://doi.org/10.61784/wjit4124

Vancouver

Zhang H. GRAPH-AWARE MULTI-TASK LEARNING FOR EARLY PREDICTION OF TIMING VIOLATIONS AND ROUTING CONGESTION IN DIGITAL IC PHYSICAL DESIGN. World Journal of Information Technology. 2024;2(1):13-20. doi:10.61784/wjit4124

BibTeX
@article{Zhang2024,
  author = {Zhang, HaiJian},
  title = {GRAPH-AWARE MULTI-TASK LEARNING FOR EARLY PREDICTION OF TIMING VIOLATIONS AND ROUTING CONGESTION IN DIGITAL IC PHYSICAL DESIGN},
  journal = {World Journal of Information Technology},
  year = {2024},
  volume = {2},
  number = {1},
  pages = {13--20},
  doi = {10.61784/wjit4124}
}
Kostenlos starten →