LLM-DSE

LLM multi-agent framework for hardware accelerator design-space exploration.

LLM-DSE uses large language model agents to search high-level synthesis directive parameters for hardware accelerators (Wang et al., 2025). Coordinated explorer/critic agents reason about performance feedback from downstream tools, outperforming classical black-box design-space-exploration methods on the HLSyn benchmark.

Code on GitHub.

References

2025

arXiv
LLM-DSE: Searching Accelerator Parameters with LLM Agents

Hanyu Wang, Xinrui Wu, Zijian Ding, and 6 more authors

arXiv preprint, 2025

Page arXiv Bib HTML
@article{wang2025llmdse, author = {Wang, Hanyu and Wu, Xinrui and Ding, Zijian and Zheng, Su and Wang, Chengyue and Prakriya, Neha and Nowatzki, Tony and Sun, Yizhou and Cong, Jason}, title = {LLM-DSE: Searching Accelerator Parameters with LLM Agents}, journal = {arXiv preprint}, volume = {arXiv:2505.12188}, year = {2025}, }