LLM-DSE: Searching Accelerator Parameters with LLM Agents
LLM-DSE is a multi-agent LLM framework (Router, Specialists, Arbitrator, Critic) for optimizing HLS directive parameters of hardware accelerators, achieving 2.55× performance gains over state-of-the-art design-space-exploration methods on HLSyn.
Hanyu Wang, Xinrui Wu, Zijian Ding, Su Zheng, Chengyue Wang, Neha Prakriya, Tony Nowatzki, Yizhou Sun, Jason Cong · arXiv 2025 · arXiv:2505.12188
Abstract
Even though high-level synthesis (HLS) tools mitigate the challenges of programming domain-specific accelerators (DSAs) by raising the abstraction level, optimizing hardware directive parameters remains a significant hurdle. Existing heuristic and learning-based methods struggle with adaptability and sample efficiency. We present LLM-DSE, a multi-agent framework designed specifically for optimizing HLS directives. Combining LLM with design space exploration (DSE), our explorer coordinates four agents: Router, Specialists, Arbitrator, and Critic. These multi-agent components interact with various tools to accelerate the optimization process. LLM-DSE leverages essential domain knowledge to identify efficient parameter combinations while maintaining adaptability through verbal learning from online interactions. Evaluations on the HLSyn dataset demonstrate that LLM-DSE achieves substantial 2.55× performance gains over state-of-the-art methods, uncovering novel designs while reducing runtime. Ablation studies validate the effectiveness and necessity of the proposed agent interactions.
Plain-language summary
High-level synthesis lets developers write accelerators in C-like code, but its performance hinges on dozens of directive parameters (pipelining, unrolling, array partitioning) whose best combination is hard to find. Classical auto-tuners treat the problem as black-box search and waste many expensive tool runs; they also cannot read the code they are tuning. LLM-DSE replaces the black box with a team of language-model agents — a Router that picks strategies, Specialists that propose directive changes, an Arbitrator that selects among proposals, and a Critic that turns tool feedback into lessons — which reason about the source code and accumulated experience in natural language while interacting with the real HLS toolchain.
Key contributions
- A multi-agent LLM architecture (Router / Specialists / Arbitrator / Critic) purpose-built for HLS directive optimization.
- Verbal online learning: the Critic converts tool feedback into reusable natural-language lessons, keeping the search adaptive without retraining.
- Tool-grounded evaluation loop connecting the agents to real HLS flows.
Key results
On the HLSyn benchmark, LLM-DSE achieves 2.55× performance gains over state-of-the-art design-space-exploration methods while reducing search runtime, and ablations show every agent role is necessary.
Resources
- Paper: arXiv:2505.12188 (PDF)
- Code: LLM-DSE on GitHub
How to cite
@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 arXiv:2505.12188},
year = {2025}
}