AgentDSE: Reasoning-Augmented Architectural Design Space Exploration

AgentDSE is a novel methodology for architectural design space exploration that integrates a general-purpose large language model (LLM) to automate the reasoning process, significantly improving efficiency. It achieves competitive design quality in deep neural network (DNN) accelerator mapping and hardware/software co-design with up to two orders of magnitude fewer simulator evaluations compared to traditional methods. This approach not only enhances the exploration process but also provides transparent decision-making through inspectable traces, offering insights into architectural hypotheses and performance bottlenecks, which is crucial for AI practitioners aiming to optimize hardware designs.