NVIDIA and Duke researchers claim up to 97% DRV solve rate with self-supervised layout model SCALE
NVIDIA and Duke University published SCALE in July 2026, a self-supervised framework that boosts AI agents' sub-2 nm DRV solve rates by 12-25%, reaching up to 97% on 100 real cases.

Researchers from NVIDIA and Duke University posted a paper in July 2026 describing SCALE, a self-supervised framework for fixing local place-and-route (P&R) design-rule violations (DRVs) at sub-2 nm nodes[1]. Tested on 100 real sub-2 nm cases, the system raised state-of-the-art AI agents' solve rates by 12-25 percentage points, reaching up to 97%[1].
Why local DRV fixing is breaking down at advanced nodes
As semiconductor manufacturing advances toward sub-2 nm nodes, local P&R DRV fixing is increasingly limited by complex rule interactions, dense multi-layer routing geometries, and foundry-specific constraints[1]. General-purpose vision-language models (VLMs) cannot fill the gap: diagnosing DRC violations from layout imagery demands precise geometric reasoning and foundry-specific rule knowledge that standard VLM training does not provide[1].
The result is a data-labelling bottleneck. Annotated layout-violation pairs are scarce, and manually tagging them at sub-2 nm is expensive and slow. SCALE is designed to sidestep that bottleneck entirely.
How SCALE generates its own training data
The framework works in two stages[1]:
- Self-supervised pre-training. Multi-layer layout geometry is serialised into structured text. A fine-tuned language model then learns to reconstruct randomly masked polygons from the surrounding back-end-of-line (BEOL) context - with no violation labels required.
- Synthetic data generation. At inference, natural-language rule constraints and high-temperature sampling steer the model toward diverse, violation-prone layout variants. An industrial signoff DRC checker validates each variant, producing DRC-annotated layout-violation pairs automatically.
Those pairs are used to fine-tune a domain-adapted DRC vision-language model (DRC-VLM). At repair time, the DRC-VLM analyses a local GDS region surrounding each violation site and proposes rule-aware geometric guidance - which is then handed off to a downstream AI agent to execute the actual fix[1].
What the benchmark shows
The paper evaluates SCALE on 100 real sub-2 nm cases covering enclosure, spacing, width, and color-spacing violations[1]. Across that set, the DRC-VLM guidance lifted agent solve rates by 12-25 percentage points, with the best result reaching 97%[1]. The authors note that general VLMs fail at the prerequisite violation-localisation task, which is why the domain-adapted model is necessary before any repair agent is invoked.
The nine-person author list spans NVIDIA Research (Santa Clara and Austin) and Duke University, with Chia-Tung Ho, Haoyu Yang, and Guanglei Zhou listed as equal contributors and corresponding authors[1]. The paper is available as arXiv preprint 2607.21850.
The immediate question for EDA tool vendors is whether the synthetic-data pipeline generalises across foundry rule decks - the paper tests one sub-2 nm technology, and rule sets differ significantly between foundries. Whether the self-supervised pre-training transfers to a new node without rerunning the full generation stage will determine how practical SCALE is as a production flow component rather than a research prototype.
Written by Electronics Insider's automated desk from the sources above and published automatically. How we work.
Related
Design & EDAETH Zürich and NYU show JEDEC test loops miss worst-case HBM2 power by 23.6%
Researchers at ETH Zürich and NYU tested 36 real HBM2 chips and found JEDEC measurement loops understate peak power by 23.6%, with read ops and data-bus toggling as the main drivers.
15 Oct 2026
Design & EDAFlowAccel's Joel Jorgensen argues chip teams must design their workflows before AI can help them hit tape-out
Joel Jorgensen of FlowAccel argues that AI tools alone won't save a tape-out schedule - teams must first design the cross-team workflows that let those tools compound into real gains.
15 Oct 2026
SemiconductorsETH Zürich and NYU find JEDEC test loops underestimate HBM2 power by up to 35.4%
Researchers at ETH Zürich and NYU tested 36 real HBM2 chips and found JEDEC-specified current loops can underestimate peak power by up to 35.4%, with implications for AI accelerator thermal design.
15 Oct 2026