Improving technical documentation for digital design : using generative AI to enhance understanding of timing diagrams

Abstract

Digital design is a complex and time-consuming process which involves translating detailed specifications and requirements into functional digital circuits. During the design process, an engineer will often deal with a wide variety of complex components. There are methods to abstract away some of the complexity but a solid understanding of the accompanying documentation is still essential. This documentation often includes detailed information about the component's operation, interfaces, timing requirements, and constraints. Understanding this documentation is crucial for seamless component integration within the larger system, preventing issues related to compatibility, performance, and functionality. Given that such component documentation can span hundreds of pages and include text, tables, images, and diagrams, exploring the possibility of increasing the interpretability of this documentation could significantly speed up development time. With recent advancements in generative AI, there is a compelling opportunity to leverage AI to assist engineers in understanding such complex documentation. In particular, this thesis is focused on designing an AI architecture capable of understanding and explaining timing diagrams. Such diagrams are complex pieces of documentation which explain the relation between signals and their timing requirements. An AI capable of understanding such complicated documentation could serve as a tool for engineers, allowing them to ask questions related to design and verification tasks. This capability will reduce the amount of time spent on analysing these diagrams and accelerate the overall development process. To this end, I developed a flexible synthetic dataset generation pipeline capable of automatically generating a large training dataset. Additionally, this work shows how this dataset can be used to fine-tune an existing model using supervised fine-tuning (SFT). Evaluation on this model shows it can identify signal values at specific time points in timing diagrams with 95.6\% accuracy. However, performance in more difficult tasks remained limited. To address these deficits, this work shows how the model can be further trained using Group Relative Policy Optimisation (GRPO). Due to resource constraints, experiments were limited to small subsets of the overall dataset, however the initial findings are promising and show there is potential in further training with GRPO to create an AI model capable of understanding timing diagrams.