AaLLM: LLM framework automates analog circuit design, 40x faster

Analog circuit design remains a notoriously manual, iterative process requiring deep expertise. AaLLM, a new open-source framework from researchers Mohammed Ayman Habib, Rylan Hart, and Morteza Fayazi, tackles this head-on with an end-to-end multi-agent LLM workflow. Unlike prior LLM-based tools that handle only sizing or topology generation in isolation, AaLLM takes user specifications and outputs a complete netlist, automating both topology generation and circuit sizing. The system automatically constructs a relevant knowledge base from research papers and textbooks, then uses a RAG (retrieval-augmented generation) model to emulate circuit design expertise without manual data collection.

The core innovation is a tri-agent feedback system: a Designer determines component values, a Critic scrutinizes them, and an Evaluator arbitrates between the two to minimize iteration overhead. This architecture dramatically improves efficiency — AaLLM achieves a 3x to 4.5x reduction in SPICE calls at inference compared to state-of-the-art multi-agent pipelines, and a 40x decrease in wall-clock time. Perhaps most importantly, the framework can generate novel topologies that achieve figures of merit (FoM) comparable to proven designs, and up to 3x higher for certain circuits. This suggests LLM-driven design can not only automate existing workflows but also explore unconventional solutions that human engineers might overlook. By making the framework open-source, the team invites the chip design community to build on these results.