Daily Arxiv

This page organizes papers related to artificial intelligence published around the world.
This page is summarized using Google Gemini and is operated on a non-profit basis.
The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.

MAHL: Multi-Agent LLM-Guided Hierarchical Chiplet Design with Adaptive Debugging

Created by
  • Haebom

Author

Jinwei Tang (Katie), Jiayin Qin (Katie), Nuo Xu (Katie), Pragnya Sudershan Nalla (Katie), Yu Cao (Katie), Yang (Katie), Zhao, Caiwen Ding

Outline

MAHL is a hierarchical LLM-based chiplet design generation framework featuring six agents that enable AI algorithm-to-hardware mapping. It includes hierarchical explanation generation, search-augmented code generation, diverseflow-based verification, and multi-granularity design space exploration. MAHL efficiently generates chiplet designs by optimizing power, performance, and area (PPA). Experimental results show that MAHL significantly improves the generation accuracy of simple RTL designs, as well as the generation accuracy of real-world chiplet designs from 0 to 0.72 at Pass@5 compared to conventional LLM. Furthermore, MAHL achieves comparable or better PPA results under specific optimization objectives compared to state-of-the-art CLARIE (expert-based).

Takeaways, Limitations

•
Takeaways:
◦
Presenting the possibility of improving the efficiency and accuracy of 2.5D integrated chiplet design using LLM.
◦
Solving complex chiplet design problems with hierarchical structures and diverse agents.
◦
Improving design space exploration capabilities to achieve PPA optimization goals.
◦
Demonstrated competitive performance compared to existing LLM and expert-based systems.
•
Limitations:
◦
It is difficult to determine whether the challenges faced by LLM-based chiplet designs (flattened design, high verification costs, and inaccurate parameter optimization) have been completely resolved based on experimental results alone.
◦
The fact that PPA results are equivalent to or superior to CLARIE only under specific optimization objectives suggests limitations in generalizability.
◦
Applicability to actual commercialization and large-scale designs requires further research.
◦
The Pass@5 evaluation method may not cover all aspects of the design and may not fully reflect the complexity of real-world chiplet designs.
👍
Made with Slashpage