Publikation
LLM-assisted Methodology for Embedded Software Performance Estimation on RISC-V
Weiyan Zhang; Muhammad Hassan; Rolf Drechsler
In: RISC-V Summit Europe. RISC-V Summit Europe, May 12-15, Paris, France, 2025.
Zusammenfassung
In this extended abstract, we present a methodology that combines a Large Language Model (LLM) with a traditional Machine Learning (ML) technique to estimate the performance of embedded software on RISC-V processors across different microarchitectures. In particular, we leverage a Retrieval-Augmented Generation (RAG)-based LLM to extract performance related information from processor specifications and source code, while utilizing the predictive capabilities of ML models to create Predictive Models (PMs) for RISC-V processors. To demonstrate the effectiveness of our hybrid approach, we present results on the performance estimation of open-source benchmarks using the generated PMs, with open-source RISC-V-based Register Transfer Level (RTL) implementations as reference models. Our results demonstrate that our proposed LLM-assisted methodology provides highly accurate predictions in comparison with the state-of-the-art methodology.