CircuitFormer: A Circuit Language Model for Analog Topology Design from Natural Language Prompt
Researchers have introduced CircuitFormer, a specialized language model designed to automate analog circuit topology design using natural language prompts. Addressing longstanding challenges in Electronic Design Automation (EDA), the team curated the largest annotated dataset of analog circuit netlists, featuring 31,341 pairs of netlists and natural language descriptions. They also developed the Circuit Tokenizer (CKT), a novel graph tokenizer that efficiently encodes netlist connectivity by mining frequent subcircuits. Unlike general-purpose tokenizers, CKT achieves constant O(1) complexity, reducing sequence length by 57% and offering superior compression with a fixed vocabulary. The resulting 511M parameter encoder-decoder transformer model demonstrates 100% syntactic correctness and an 83% functional success rate across major analog circuit categories. CircuitFormer significantly outperforms state-of-the-art open-source large language models while requiring 240 times fewer parameters. This advancement bridges the gap between natural language processing and hardware design, offering a scalable and efficient solution for automated analog circuit generation. The associated dataset has been made publicly available to support further research in this domain.
Wire timeline
CircuitFormer: A Circuit Language Model for Analog Topology Design from Natural Language Prompt
Researchers have introduced CircuitFormer, a specialized language model designed to automate analog circuit topology design using natural language prompts. Addressing longstanding challenges in Electronic Design Automation (EDA), the team curated the largest annotated dataset of analog circuit netlists, featuring 31,341 pairs of netlists and natural language descriptions. They also developed the Circuit Tokenizer (CKT), a novel graph tokenizer that efficiently encodes netlist connectivity by mining frequent subcircuits. Unlike general-purpose tokenizers, CKT achieves constant O(1) complexity, reducing sequence length by 57% and offering superior compression with a fixed vocabulary. The resulting 511M parameter encoder-decoder transformer model demonstrates 100% syntactic correctness and an 83% functional success rate across major analog circuit categories. CircuitFormer significantly outperforms state-of-the-art open-source large language models while requiring 240 times fewer parameters. This advancement bridges the gap between natural language processing and hardware design, offering a scalable and efficient solution for automated analog circuit generation. The associated dataset has been made publicly available to support further research in this domain.
cs.AI updates on arXiv.org