Challenges of using AI in EDA

The data validity and consistency issues in AI needs to be considered while using AI in EDA. Can we use ChatGPT, in the chip design process? 

The immediate answer is No because the Engineers just cannot use RTL which was written before. 

So, for GenAI, the question is when AI will be able to generate Verilog/VHDL or for example, replace constrained random test pattern generation? OpenAI, probably takes billions of tokens to train to acquire desired intelligence to generate. 

Even, after AI generate RTL, the next step will be how to optimise it? 

The drawback with using large language model(LLM) in EDA is that the designers may need to spend hours in debugging poorly written RTL. Which they could probably avoid by making the design afresh. 

There will be another challenge with GenAI, while modelling designs with AI in the loop, will the code be safe and secure or going to have challenges in IP reuse?

The biggest challenge for AI adoption in EDA is how to get data that spans entire industry to properly train a powerful  model. Generative EDA, may have to wait a little longer. 

AI will be used in design where it is checked by humans and gives humans a template to start with, so they can move forward more effectively, more quickly.

While EDA is adopting machine learning and other components of AI, it is not probably possible to throw away many of the existing algorithms immediately. 

Those algorithms help to accelerate things, and to automate processes. However, for those tasks machine learning can help. 

Sometimes machine learning can improve the primitive algorithms and make them more accurate. The EDA industry wants accuracy. Eventually, the models will be trained well enough where we’ll get that surety.