AaLLM: An End-to-EndAnalog Circuit Design Framework from Topology Generation to Sizing UsingLargeLanguage Models

Analog circuit design is a time-consuming, iterative process in a nonlinear and high-dimensional design space that relies heavily on expert intuition. This paper presents AaLLM, an open-source end-to-end multi-agent large language model (LLM) workflow facilitating both topology generation and circuit sizing based on user specifications. The methodology involves automating the creation of a knowledge base from various research sources, employing a Retrieval Augmented Generation (RAG) model to emulate design expertise. Additionally, a tri-agent feedback system comprising a Designer, Critic, and Evaluator minimizes circuit sizing iterations. AaLLM leverages a fine-tuned Sequence-to-Sequence (Seq2Seq) model for generating novel topologies by learning and recombining conventional circuit connections. Results demonstrate that AaLLM-generated topologies achieve a figure of merit (FoM) comparable to existing designs, with up to three times higher performance for select circuits. Testing demonstrates a significant reduction in the number of SPICE calls, with a decrease of 3 to 4.5 times compared to state-of-the-art multi-agent LLM pipelines and a 40-fold reduction in wall-clock time over existing methods.