IF-THEN-AI: Improving Feedback of Tooling for Hardware Engineering using Generative AI

Achieving high productivity with hardware development toolchains is difficult. When issues arise, understanding the available feedback (simulation and synthesis error messages, waveforms, output files) to determine the root cause is challenging. This motivates using machine learning (ML) to assist in this area. We propose utilizing Large Language Models (LLMs), which have demonstrated significant capabilities for comprehending and producing complex structured and unstructured text. Our goal is to train/fine-tune LLMs over the developer workflow itself, starting from open-source tools (e.g. iverilog, Yosys) and their usage scripts. These would be combined with relevant design context such as files and simulation data, including logs and waveforms. Some data may be project-specific, while others may be more regularly structured (such as error messages).

We will work to the following goal: after an issue is encountered, the LLMs will be able to provide rectifying guidance, thus accelerating developer productivity. As solely advisory models, these LLMs would not introduce the same level of risk as those used for code writing. To ensure that there would be no costly infrastructure requirements, we would focus on smaller and locally-deployable models.