Evaluating Large Language Models for High-Level Synthesis
Ali Rizvi, Nia Simon, Janet Tocho, Asil Yongaci, Stefan Abi-Karam, Cong Hao · 2024
Large language models (LLMs) are comprehensive tools, that are capable of modeling natural languages, such as human reasoning, and code structures, and even more recently, using expansive collective databases are even now able to create generated information based on a given user prompts. Given the extensive capabilities of LLMs and their diverse knowledge in many fields, this paper explores the use of LLMs for hardware design. Specifically improving LLM’s capability for code editing and generation of C++ code targeting high level synthesis (HLS) for FPGA design, as well as creating generated code based on a programmable prompt. We built a frame work to interact with LLM tools. Results indicate that there is success in creating an external framework for zero-shot code editing, indicating that the approach is valid with minor model performance. Future work looks to improve the framework’s accuracy for the current model and expand usage to other models as well as code generation.