Never miss a story from Artificial Intelligence (AI). Subscribe for in-depth analysis and articles.
I have always relished technology discussions and expert opinions on how technology will unfold and shape the future. In the panel discussion during CadenceLIVE Europe, experts shared their thoughts about the impact of the AI revolution on EDA. They also discussed the benefits of semiconductor design, its impact on academia, risks, and challenges. In my previous post, I covered the panel's opinions on the current state of Generative AI and its potential to revolutionize the EDA industry.
Moving ahead to the next question from moderator Rosa.
Can generative AI improve the speed of learning?
Hussam: The use of generative AI can increase the learning rate of students, enabling them to become experts in their field in less time. Large language models generated by AI can help individuals who are already educated to learn faster, which is especially beneficial when it comes to finding RTL bugs. However, he expressed concerns that generative AI may create difficult-to-find bugs, which could be disruptive in industries such as healthcare and automotive that rely on chips. Therefore, we need to exercise caution when relying on generative AI, since it learns from vast amounts of unseen data and may create challenging bugs that are impossible to resolve. We do not want to risk building chips with undiscovered bugs that could cause problems in the future.
Rosa highlighted that as chips become more complex and technology shrinks, the cost of verification rises. She raised the following question pertinent to all the chip manufacturers.
Can generative AI reduce the rising verification cost?
Jean Christophe: Certainly! In general, the development process is often the most extensive phase, as the complexity of each generation tends to increase. For example, we now have more variations and configurations to consider, which means bias has a significant potential to impact our verification efforts. One area of interest and challenge is using technology to test and verify these configurations more efficiently. Finding a satisfactory solution to this challenge would save time and resources in the long run.
Rod: AI techniques can help to find the failures at the earliest. Using AI techniques in Cadence tools has brought immediate benefits for the partners and has made the verification cycle more productive. There are some exciting ideas in the pipeline as well. For instance, currently, interpreting the high-level spec and devising a verification plan is laborious and manual. Large language models hold enormous potential to assist this task. While they may not be able to complete the process entirely, they could take us 75% of the way there, leaving the rest to be done manually. This could introduce a significant level of automation to the process that could help us.
Hammond: Recent advancements in language models have led to their enhanced ability to understand unstructured specifications and contextual information related to a particular task. These models are now capable of learning how to use external tools as well. By training the model to call upon external programs, it no longer needs to be the sole source of information and can learn to use other resources. This development is particularly intriguing in Education Technology (EdTech), where many tasks rely on tool collections. Language models can serve as a new interface where designers can assign high-level tasks, and the model can break them down into smaller steps involving verification and context analysis. Additionally, the model can call external tools to help complete those steps.
What are the challenges with the usage of Generative AI on Chip design?
The first thing that comes to mind for many people is the issue of intellectual property. What is your opinion on this matter? Rosa asked Hammond.
Hammond believes this bottle is already open, and Gennie is out of it. ChatGPT has gained a significant amount of insight from the data available to them without anyone providing it. However, it is still under consideration whether they had the right to do so initially. Protecting intellectual property is challenging, particularly in the face of globalization. With collective efforts, it becomes easier to safeguard the rights of data generators. The industry might need to come together to find a solution. Everyone should benefit from the models created while also respecting the rights of those who generate the data. The critical challenge will be to figure out who owns the data and can use the created models. It is essential to strike a balance between the learning rate and the rights of data generators.
Rod: It is widely recognized that the issue of ownership in the public domain is a complex matter that requires careful consideration. In the commercial domain, particularly about design collateral, the question of who has ownership over the material created and used is a major concern. This is a matter that not only the education sector but also the legal community needs to address. It is necessary to establish how to measure ownership and assign it appropriately. It is not practical for everyone to create their own models, and thus, a base model is needed to supply the necessary human interaction and other essential elements. As a member of the education industry, we will contribute to this model and add our proprietary data. However, creating individual models for two trillion properties is not workable. Therefore, we need to establish the legal parameters of the data we train on, particularly in the public domain. Although we are aware of the source and owner of the data we use, the situation with data in the public domain is much more complex.
How do you see that from a business perspective?
Jean Christophe acknowledges that there is a choice between waiting for countries worldwide to change their laws and recognize ownership rights or continuing to work on it. He mentioned that there are limitations in terms of what can be done and that the implementation of these changes in production requires waiting to see how it will all turn out. Further, he expects that it will take at least another year before any considerable progress is made in this area worldwide.
Hammond added humor to this question, saying this is primarily a matter for lawyers and politicians rather than engineers. While tools like Co-Pilot may offer some confidence, they are not the ultimate decision-makers.
To generate more awareness amongst the next generation of students about the benefits and the risks of generative AI, Cadence, Arm, TU Munich, UNSW Sydney, University of Southampton, and Europractice announced a GPT Design Contest, where participants have 24 hours to create a design, which corresponds to a specification and will be verified using Cadence Verification IP. The design contest will be held during the DATE 2024 conference.
If you missed the chance to attend this panel discussion at CadenceLIVE Europe 2023, do not worry. On-demand videos are coming soon. You can register at the CadenceLIVE On-Demand site to watch it and all other presentations.