TL;DR
This paper introduces StructRTL, a graph-based learning framework that leverages control data flow graphs and knowledge distillation to improve RTL design quality estimation, outperforming previous methods.
Contribution
It presents a novel structure-aware graph self-supervised learning approach for RTL quality estimation, integrating cross-stage knowledge transfer for enhanced accuracy.
Findings
Outperforms prior RTL quality estimation methods
Establishes new state-of-the-art results
Demonstrates effectiveness of structural learning and knowledge distillation
Abstract
Estimating the quality of register transfer level (RTL) designs is crucial in the electronic design automation (EDA) workflow, as it enables instant feedback on key metrics like area and delay without the need for time-consuming logic synthesis. While recent approaches have leveraged large language models (LLMs) to derive embeddings from RTL code and achieved promising results, they overlook the structural semantics essential for accurate quality estimation. In contrast, the control data flow graph (CDFG) view exposes the design's structural characteristics more explicitly, offering richer cues for representation learning. In this work, we introduce a novel structure-aware graph self-supervised learning framework, StructRTL, for improved RTL design quality estimation. By learning structure-informed representations from CDFGs, our method significantly outperforms prior art on various…
Peer Reviews
Decision·Submitted to ICLR 2026
1. Correctly predicting design metrics at an earlier stage in the design cycle is very useful since it can reduce the overall design cycle time. 2. Using CDFGs is well motivated. 3. The paper is generally well-written and easy to follow.
1. while larger than some prior work, however the current dataset size is still small. 2. 80% of the designs have less than 600 nodes which raises concerns about whether the method scales to real life designs and whether the current dataset truly represents real-world complexity. 3. Proposed knowledge distillation requires running synthesis to get the netlists. Scalability will become an issue for big designs.
a. State-of-the-Art Performance: Clear and substantial outperformance across multiple quality metrics (area R²=0.8676, delay R²=0.8872) on a large, modern dataset, providing strong evidence for the structure-aware approach's superiority. b. Rigorous Experimental Validation: Comprehensive ablation studies systematically validate each architectural component (GNN backbone, pretraining tasks, positional embeddings), proving their individual contributions are critical and non-redundant. c. Effecti
a. Limited Generalization Assessment: All experiments confined to OpenABC-D dataset; no evaluation on other RTL benchmarks or real-world proprietary designs raises questions about performance on diverse circuits outside the training distribution. b. Insufficient Knowledge Distillation Analysis: The source of the teacher model's superiority isn't deeply explored; missing ablation comparing netlist-trained vs CDFG-trained teachers to validate claims about "low-level insights." c. Indirect LLM Co
* Very strong experimental results. * Pretraining tasks are sounds and essential for representation learning, leading to the strong downstream task result of overall quality estimation.
* This work implements various ideas well but is ultimately heavily inspired by prior work. Graph representation (GraphMAE, MaskGAE) and fine-grained knowledge distillation (VeriDistill).
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
