Generative AI for agile hardware development

Abstract

The integration of generative Artificial Intelligence (AI) into microprocessor design marks a paradigm shift in hardware development. The CyberRio project at RIOS is among one of the earliest attempts of using a Large Language Model (LLM) to create a RISC-V CPU design from scratch. Leveraging the cutting-edge GPT- 4, CyberRio deploys AI generated content across the entire CPU design lifecycle, from initial design specifications, RTL implementations to functional verifications. We believe that domain-adapted generative AI models can be highly effective in the field of chip designs. We are also developing an LLM based framework called Rosemary with various chip design domain knowledge adaptation techniques in tokenization, continued pretraining, model alignment, and retrieval models. Our goal is to create an engineering AI chat assistant that can support engineers in development, a document generator that automatically generates modifiable drafts of technical documentation, a smart tool that can summarize and analyze bugs, as well as generate EDA scripts. Through our explorations with generative AI, we have found that domain-adaptive continued training and fine-tuning models can perform better domain specific tasks than the vanilla LLaMA2 baseline model. This gives us confidence in the potential of domainspecific enhanced LLMs like Rosemary can be more effective in the future agile hardware development.