What Does Semiconductor Disruption Look Like?

When conducting interviews for my article on the incorporation of AI within EDA tools, Anand Thiruvengadam, senior director and head of AI product management at Synopsys, said, “AI has the potential to transform how customers do chip design. The entire EDA flow can be disrupted with AI.” He is not alone in making this kind of statement. Each year, I do a predictions piece, and I ask about how AI will disrupt EDA. Quite disappointingly, nobody has ever given me a good answer beyond their recent press releases.

Disruption is very difficult to foresee – I understand that, but the best people seem to be able to come up with is that agents will make people more productive. I do not believe that is disruption. That is a productivity aid that enables an engineer to do what they do already better or faster. Don’t get me wrong – this is valuable, but not disruptive. It enables companies to potentially replace several engineers with a single engineer and a bunch of agents. They now have the collective wisdom that has been captured within those agents. But true disruption happens when the task they perform changes.

When ChatGPT made its public appearance, almost three years ago, I said at a DAC panel that it signaled the end of the term “to Google” something. People thought I was crazy, and yet in just three years, who types something into a Google search bar and just expects to see a list of websites? Probably nobody. At the very least, the search engine has scraped the relevant websites and provided a summary. That often provides enough information to satisfy the reason why you did the search. In a recent opinion piece in the Communications of the ACM, Moshe Vardi went further and said it potentially signaled the entire collapse of the World Wide Web, including the notion of websites – because who will spend time and effort building them if nobody ever actually visits the site? All you are doing is providing information on which AI can be trained, and good luck trying to monetize that.

Another part of his thesis is that as AI starts to eat its own tail and thinks it is learning from itself, bias and errors become increasingly built into the system. In this context, it can cement bias like racism, sexism, etc. Some of those issues relate to what may potentially happen in chip design and EDA. If we do not start to discuss them, a lot of costly mistakes could be made.

Chip design is complex. Hugely complex (my high school English teacher would have rapped my knuckles for that). The increase in complexity, driven by Moore’s law, has meant that compromises had to be made. Most of that was in the area of creativity. It was quicker, cheaper, safer to incrementally add onto what was already known rather than go back to a blank sheet of paper each time. The introduction of IP cemented this notion even further, and the need to preserve legacy software locked in hardware architectures. This is all bias in terms of design.

The rise of parallel processing did not come about because it was deemed to be good, it came about because the limits of single processor architectures had been reached. It took a further decade or more before it started to see significant adoption, and the introduction of early machine learning technology became the compelling event for its serious development. Assuming that an AI-powered chip design tool existed and was trained on the collective wisdom of the entire industry at that time, would it have been able to make that leap on its own? I seriously doubt it.

While it would have known about parallel processing (which was a lot more common back in the 1980s than it was in the 2010s) and it would have known how to write code to target them, the bias of single processor designs would overwhelm. It would have learned about using multiple single processors, often acting as agents to contribute little bits to the function of the total (think audio processor, USB controller, graphics controller…) rather than a central heterogeneous processor that could handle everything more efficiently.

Did Nvidia set out to create an AI processor? No. They made incremental steps to satisfy the needs of their customers. It was when one application, computer vision, which had a sufficient dataset for training, plus the availability of suitable hardware, that deep learning was first shown to be superior to rule-based systems. That sparked many of the advancements we see today.

The notion of a semiconductor AI system being trained on all existing data is never going to happen. Each company focuses on a particular type of chip, a particular type of problem, such as mobile, automotive, data center, etc. Their bias is even more ingrained. Few hardware companies are also developers of software, even though they may need some hardware-dependent software. AI is not going to be told to invent something radically new, it will be asked to make incremental improvements to either that hardware or the software.

What could semiconductor disruption look like? I don’t think it will happen overnight and I don’t think it will all happen at once. I would expect EDA disruption to first appear in something like high-level synthesis (HLS), where a tool could be trained on a large collection of architectures. A suitable amount of data exists for this training, and it could be enhanced by individual companies. This would enable it to take in an English-like specification and generate code that would then be taken through a traditional EDA flow.

While that has been the goal of HLS since its inception, it has proven to be too difficult or only viable in constrained applications. SystemC has not really been accepted as an input language, and even then, tools require a highly constrained set of language constructs. But if AI can help make this a reality, the user base could expand 10X to 100X when custom design becomes accessible to a much larger community. That is enough to shift the needle for the entire EDA flow and was why there was such a large investment in this technology when it first emerged.

Virtual prototypes will also need to be able to operate on those specifications so that errors or omissions can be identified, and sequential equivalence checking will need to be significantly advanced so that confidence can be shown in the AI transformations. Over time, that new core flow will bring in more of the agentic helpers to look after things like power, cost, etc.

Disruption happens when something changes, not when something becomes optimized.