LLM-DSE: Searching Accelerator Parameters with LLM Agents

TL;DR

This work introduces LLM-DSE, a multi-agent framework that uses specialized LLM-driven agents (Router, Specialists, Arbitrator, Critic) to optimize HLS directives for domain-specific accelerators. By framing design-space exploration as a cooperative tree search with a dual-stage filtering pipeline and context-management tools, it achieves substantial speedups over heuristic and model-based baselines and generalizes to multiple toolchains and larger programs. Ablation studies confirm that agent interactions and LLM-based guidance are essential for performance, while token-efficiency considerations illustrate practical deployment costs. Overall, LLM-DSE offers a scalable, domain-aware method to improve DSA performance, enabling faster design-turnaround and broader accessibility of HLS-based acceleration development.

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 $2.55\times$ 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. Our code is open-sourced here: https://github.com/Nozidoali/LLM-DSE.

Figures (9)

• Figure 1: Searching accelerator parameters is a challenging task.
• Figure 2: LLM-DSE overview. The inputs are the HLS code with a set of parameters to be determined, and the design space that specifies the value ranges. The output is a design with all parameter values. It coordinates different agents and navigates in the design space via a collaborative process between the router, a group of specialists, an arbitrator, and the critic.
• Figure 3: The Merlin Compiler simplifies hardware development by allowing designers to specify hardware directive parameters ("#pragma ACCEL") within a given template. Still, due to the interdependent nature of hardware directives and the vast combinatorial space of possible parameters, identifying an efficient parameter set remains a challenging task, even for experienced designers.
• Figure 4: Different combination of HLS parameters means different micro-architecture design. Here, each processing element (PE) multiply $a[j]$ with $b[i]$. When different directive parameters are selected, the underlying hardware will also be different, resulting in different latency and resource consumption.
• Figure 5: Router's prompt.