User:Wenji838/Circuit Foundation Model

Circuit Foundation Models (CFMs)

Introduction

‘’‘Circuit Foundation Models (CFMs)’’’ are an emerging class of artificial intelligence (AI) models designed specifically for VLSI circuit design and EDA. Unlike traditional AI techniques tailored to individual tasks, CFMs use a two-stage training process consisting of self-supervised pre-training on large unlabeled datasets, followed by efficient fine-tuning for specific tasks.

Definition and Paradigms

CFMs are characterized by:

Generalization across tasks

Reduced reliance on labeled circuit data

Efficient adaptation to new tasks

Advanced generative capabilities

CFMs are categorized into two main types:

Encoder-Based CFMs

Encoder-based CFMs primarily encode circuit designs into embeddings, capturing intrinsic properties for predictive tasks such as performance, power, area estimation, and functional verification.

Examples include:

HARP

ProgSG

Design2Vec

DeepGate family

Decoder-Based CFMs

Decoder-based CFMs use Large Language Models (LLMs) to perform generative tasks such as HDL code generation, design verification, debugging, and optimization.

Examples include:

RTLLM

ChipGPT

VerilogEval

Historical Context and Development

CFMs emerged after the success of foundational AI models like BERT and GPT in NLP and CV. CFMs started gaining significant traction around 2022 due to their transformative potential in circuit design.

Techniques and Applications

CFMs integrate various AI techniques:

Self-supervised learning (contrastive learning, mask-reconstruction)

Supervised pre-training

Multimodal learning (integration of textual, structural, and layout modalities)

CFMs have applications in:

Early-stage design quality evaluation

Functional verification and debugging

Circuit generation and optimization

Design space exploration

Security verification

Challenges and Future Directions

Despite rapid advancements, CFMs face several challenges:

Limited availability of open-source circuit data

Scalability and performance issues for large circuit datasets

Integration and alignment of multimodal data across design stages

Future research directions include enhancing scalability, developing synthetic circuit data for training, and unifying encoder-decoder architectures for improved functionality and efficiency.

Key Contributions and Impact

CFMs significantly reduce design effort, cost, and turnaround time in VLSI circuit design by automating and streamlining the design process. Their ability to generalize and generate novel circuit solutions positions them as crucial tools in future EDA methodologies.

References

Fang, W., et al. (2025). ‘‘A Survey of Circuit Foundation Model: Foundation AI Models for VLSI Circuit Design and EDA’’. arXiv:2504.03711.