Iteration And Hallucination

Iteration loops have been a vital aspect of EDA flows for decades. Ever since gate delays and wire delays became comparable, it became necessary to find out if the result of a given logic synthesis run would yield acceptable timing. Over the years this problem became worse because one decision can affect many others. The ramifications of a decision may not have been obvious to an individual tool in the flow. Running tools serially led to big problems, and in the case of logic synthesis, it became almost impossible to close timing, even with iteration, because there was no concept of learning in the loop.

These multi-factor dependencies were solved by closely linking decision-making tools, estimators, and checkers into single tools, so that quick checks can be made along with decisions to find out the likelihood that a given choice may end with a bad end result. We are seeing increasing areas where this is necessary, and many of them have interactive feedback loops so that users can guide the tools based on the expert knowledge they have.

AI has a highly publicized ability to hallucinate. Many of these hallucinations are funny when not in something critical, especially when it comes to generating graphics. One of my hobbies is modeling British railways from around 1950. A while back I tried to use AI to generate some backscenes. AI fundamentally does not understand the concept of a train as having two-flanged wheels that run within a pair of tracks. Some of the images generated had an arbitrary number of rails and it got highly comical when turnouts were included. Rails went everywhere, some vanishing into thin air, and some trains levitated over them. See how many errors you can detect in the image below.

Image created by DALL·E 3 when asked for a British railway junction from the 1950s.

The other day I was reading Communications of the ACM (May 2025), which had an article by Samel Greengard titled, “Shining a Light on AI Hallucinations.” In it he said, “As words become numeric vectors and numerical values multiply by an order of magnitude inside a model, things get fuzzy and meanings get blurred. The model simply regurgitates words based on probability. ‘AI hallucinations are a feature rather than a bug,’ said Anima Anandkumar, Bren Professor of Computing and Mathematical Sciences at the California Institute of Technology (Caltech). ‘Models are incentivized to produce plausible text rather than factual text.’”

This is why many people are looking to reign in these hallucinations using retrieval augmented generation (RAG). The purpose of these is to fact check what is being generated and to make sure that what is generated is actually plausible. This is beginning to sound very similar to what we have been doing in EDA for a long time – not that we believe that EDA tools hallucinated. They are just not smart enough to know something wasn’t going to work, or that an engineer was rather too quick to finish something on a Friday afternoon so they could get an early start on the weekend.

Another person said to me that when they used AI to help them generate an analog circuit, it came up with a completely new architecture they had ever tried before. I am not sure if someone else had, and then AI had been trained on that alternative architecture, but we could argue whether that was a smart decision made by the tool or a hallucination that turned out to be brilliant. Within EDA we have learned that everything has to be verified and thus it would appear that we have a leg up on many other applications of AI. We are so accustomed to verifying everything to the maximum ability possible in the timeframe and cost budget, and we have a lot of experience (far from perfect) about setting priorities on those resources.

It may require some rejiggering of verification flows, but if it could be done successfully, it could be the RAG for GenAI being used in the EDA development flow. The big stumbling block is that when these techniques have been used within a tool, they have relied on fast estimators, and functional verification is certainly not fast. As an industry, we desperately need new abstractions for functionality that can give us a quick sense of the validity of decisions that are being made, and not always think of it as requiring a full regression run.

We have made progress in using AI to look at the stimulus sets that may provide the maximum feedback for a minimum cost, but we should also be able to ascertain the necessary abstraction to validate an idea to a reasonable degree. That is something we don’t know how to do today, even though it has been going on for many of the other multi-physics problems within the wider electronics domain. Digital twins and reduced order models are examples of this.

In a recent story about mixed-signal systems, several people said that it was important to have model generators that could spit out the required abstraction for the verification of defined aspects of a design. We need that for all forms of functionality, and it can be done. One example is Arm, which has long had a single golden model for a processor, from which several other models are generated. I am sure it takes longer to generate that master model to begin with, but then we can be certain that all models are completely aligned with each other, and there is no need to verify every model that is independently created.

With the right verification in place, we could allow AI to hallucinate as much as it wants, and maybe it will provide breakthroughs that would never have come about otherwise. Given that we have never questioned decisions that were made 50 years ago, AI could provide the industry with a much-needed re-architecting phase, and that could bring about huge performance and power gains. We certainly need it.