Published on October 30, 2023, 1:10 pm
In a research paper published today, Nvidia semiconductor engineers unveiled how generative artificial intelligence (AI) can revolutionize the intricate process of designing semiconductors. The study showcased how large language models (LLMs) trained on internal data can create assistants that enhance productivity in specialized industries.
Semiconductor design is an exceptionally complex undertaking, involving the meticulous construction of chips with billions of transistors on 3D circuitry maps thinner than a human hair. It requires coordination among multiple engineering teams spanning several years, each specializing in different aspects of chip design and employing specific methods, software programs, and computer languages.
The lead author of the research paper, Mark Ren, who is also an Nvidia Research director, believes that large language models will play a crucial role in improving processes across the board. Bill Dally, Nvidia’s chief scientist, announced the paper during a keynote speech at the International Conference on Computer-Aided Design held in San Francisco. Dally described this effort as an important first step in leveraging LLMs for the complex task of designing semiconductors.
Nvidia’s research team developed a customized LLM called ChipNeMo, which was trained on internal data to generate and optimize software while assisting human designers. The ultimate objective is to apply generative AI to every stage of chip design to significantly enhance overall productivity. As part of their exploration, the team has already developed use cases such as a chatbot, code generator, and analysis tool.
The most well-received use case so far is an automated analysis tool that streamlines the time-consuming task of maintaining updated bug descriptions. Additionally, there is an ongoing development of a prototype chatbot that helps engineers quickly find technical documents and a code generator that creates snippets of specialized software for chip designs.
The research paper primarily focuses on the team’s efforts to gather design data and create a unique generative AI model. This process can be applied across various industries. Starting with a foundational model, the team used Nvidia NeMo, a framework for building and customizing generative AI models, to refine their model. The final ChipNeMo model, with 43 billion parameters and trained on over a trillion tokens, demonstrated its ability to discern patterns effectively.
This study serves as an exemplary demonstration of how a highly technical team can improve a pretrained model using their own data. It emphasizes the importance of customizing LLMs as even models with fewer parameters can achieve or surpass the performance of larger general-purpose LLMs. Careful data collection and cleaning are crucial during the training process, and users are advised to stay updated on the latest tools that simplify and expedite their work.
The semiconductor industry is only scratching the surface when it comes to harnessing the potential of generative AI. The findings from this research provide valuable insights for enterprises interested in developing their own customized LLMs. Nvidia recommends utilizing the NeMo framework available on GitHub and the Nvidia NGC catalog.
In conclusion, Nvidia’s research paper highlights how generative AI can revolutionize semiconductor design by leveraging large language models trained on internal data. The introduction of specialized AI models holds significant promise for enhancing productivity in highly specialized industries like semiconductor design. The ongoing development of use cases such as chatbots, code generators, and analysis tools demonstrates the practical applications of generative AI in chip design.