MACO: A Multi-Agent LLM-Based Hardware/Software Co-Design Framework for CGRAs

TL;DR

The paper presents MACO, an open-source multi-agent LLM-based framework for CGRA HW/SW co-design that iteratively explores the design space, repairs design errors with rule-based mechanisms, and selects high-quality designs using adaptive confidence and self-learning. By combining exponentially decaying exploration, repair libraries, and a two-tier evaluation loop, MACO reduces design turnaround time and improves PPA compared with baselines. Experimental results across multiple kernels and domains show substantial power reductions and performance gains, with MACO designs validated through ASIC flows. This approach demonstrates the viability and practicality of LLM-driven automation for complex CGRA design tasks.

Abstract

Coarse-Grained Reconfigurable Arrays (CGRAs) offer high performance and energy efficiency across domains, yet design remains difficult due to a vast, interdependent space and costly manual iteration. We present MACO, an open-source multi-agent LLM framework for CGRA HW/SW co-design. MACO generates and refines architectures through four stages: HW/SW Co-design, Design Error Correction, Best-Design Selection, and Evaluation & Feedback, iteratively improves power, performance, and area (PPA) via agent reasoning and closed-loop feedback. To traverse the space efficiently, we introduce exponentially decaying exploration; to accelerate convergence, we incorporate an LLM self-learning mechanism that adaptively selects promising candidate CGRAs. In addition, we propose a rule-based mechanism to correct CGRA design errors. Evaluated against other state-of-the-art methods, MACO achieves superior PPA while substantially reducing human effort, highlighting the promise of LLM-driven automation for practical CGRA design.

Figures (8)

• Figure 1: A Typical CGRA Design with Compiler and Architecture Components.
• Figure 2: An Overview of the MACO Framework for CGRA Design and Details for Stages 1–3.
• Figure 3: Prompt A and B to guide LLM for balancing the exploration and exploitation.
• Figure 4: The implementation of CGRA design selection with LLM self-learning based on comparison of results selected by the EDA tools and the LLM.
• Figure 5: Normalized Power efficiency (performance per watt) for typical kernels in three domains by (a) three models used in MACO. (b) different methods.(normalized to the per-kernel maximum value)