UCF DRACO LAB Design of Resilient Architectures for Computing

Automated RTL Generation

Primary Research Question: What capabilities to current code synthesis tools have for developing RTL code?

Automated synthesis tools have made large strides in recent years in their capabilities to produce production grade code with minimal specification. GPT systems like Bard and ChatGPT have established a new state-of-the-art within the program synthesis field, but their applications to Verilog and other HDLs is an open question. Initial experiments have demonstrated significant restrictions on the off-the-shelf adoption of these tools, and thus this project seeks to answer questions such as: what problems have transformer architectures ingested sufficient training data to solve properly within the verilog space? what capabilities have these systems developed for evaluating themselves? and what interaction paradigms are most effective for interfacing with these novel tools? This project seeks to evaluate a number of AI models in their code synthesis capabilities and, using an open coding technique, describe the quality of the code produced by these systems in comparison to wholly human generated code.