Michelle Thompson W5NYV and Matthew Wishek NB0X presented about the Role of Artificial Intelligence and Machine Learning (AI/ML) in Register Transfer Logic Design (RTL) Generation at a joint meeting of Open Source Digital Radio IEEE Local Group and the San Diego Chapter of the IEEE Information Theory Society. The meeting was held on 28 January 2025 and a recording can be found at https://youtu.be/8xDxeUxWTCc
AI/ML in RTL Design Generation Meeting Summary
In the meeting, Michelle and Matthew presented about the potential of artificial intelligence and machine learning in Electronic Design Automation (EDA) frameworks and the RTL design generation process. They identified core concepts from an open-source perspective and discussed the importance of reducing schedule, technical, and cost risks in the design process. Matthew highlighted the iterative nature of the design process and the need for early identification of design deficiencies and incorrect assumptions. He also suggested the use of AI agents to guide the high-level synthesis process, assist in design space exploration, and improve the performance of place and route.
Michelle discussed the potential of open-source EDA and recommendations from a European white paper, highlighting the challenges of limited access to EDA software and the workforce shortage. She mentioned the positive impact of open-source initiatives like the RISC-V processor and the Google Skywater Process Design Kit (PDK). The techincal recommendations included focusing more on multiple areas including analog and mixed-signal designs, interoperability and verification, and system-on-chip integration. She also emphasized the importance of proper licensing, funding, sustainability, and industry training for both open-source and AI/ML projects and highlighted the recommendation from Free and Open Source Silicon (FOSSi) Foundation for more conferences and events to better understand and filter the vast amount of information in the field.
Michelle discussed the use of large language models and automated writing of HDL for design IP. She shared her experience with ORI’s Remote Labs Matlab HDL Coder toolbox, which can produce high-quality, human-readable HDL code but requires a lengthy process and is not suitable for complex monolithic designs. She also mentioned the use of AI and ML in deep learning hardware models and their potential to inform the design process. Michelle also presented a survey of experts in the field, which showed a bias towards high-quality from proprietary tools and lower quality expectations from open source and AI/ML tools. Matthew then introduced a set of relevant papers he had found in his literature search, focusing on areas such as RTL generation, hardware verification, testbench generation, and formal verification. He emphasized the potential of these papers to inform and improve the design process.
Matthew presented about the role of of AI/ML in analog design, signal processing, and network planning. He suggested that AI tools can help with analog design, potentially enabling non-specialists to experiment and learn. He discussed the use of AI/ML in detecting convolutional codes, channel estimation, and radio map generation for network planning. There is significant overlap with information theory in these applications, particularly in tracking entropy levels in codes and signals. Michelle mentioned the potential for presentations on this topic at the Information Theory and Applications Conference 9-14 February 2025. Michelle also highlighted successful technology demonstrations by DARPA in 2019 showing improved spectrum efficiency using AI/ML models.
Daniel, Michelle, and Matthew discussed the pros and cons of using proprietary versus open-source Electronic Design Automation (EDA) tools. Michelle described the challenges of open-source projects, such as the lack of funding, limited market size, and difficulty in maintaining quality. Matthew shared his experience with open-source simulators, noting their limitations in terms of feature set parity and language coverage. Daniel suggested that open-source solutions might take longer to develop but could benefit from advancements in AI tools. Michelle and Matthew agreed that the EDA market is small, which contributes to the high cost of proprietary tools. The consensus from the participants was that while there are good open-source projects addressing AI/ML in RTL design generation, many of them may not yet be ready for complex designs.
Future Work for ORI AI/ML Team
It was resolved to pick one or more of the many tools and projects mentioned in the presentation, and give it a try. Reports about the first-hand experiences would then be shared in future IEEE meeting collaborations, to put the recommendations from FOSSi and the rubrics from Matthew into better context.