Hallucination And Innovation At DAC

At DAC this year, I had the pleasure of moderating an intimate chat between Alon Shtepel, senior director for ASIC at Micron, and Abhi Kolpekwar, vice president and general manager for digital verification technology at Siemens EDA. The assigned topic was generative AI in design and verification, with the more provocative subtitle asking if we are hallucinating or innovating?

L-R: Brian Bailey, Alon Shtepel, Abhi Kolpekwar.

With both of them having a background in verification, I started by asking if we should invest resources in finding bugs through verification or improving the development process so that bugs do not get created in the first place.

Abhi expressed his view that because of the shortage of resources in the industry, designers will be expected to do an increasing amount of verification. “We will basically converge them, but you cannot eliminate the need for verification. Correct-by-construction is an approach that many would like to see, but this is not going to completely eliminate verification. Both are going to be important, and both will have a role moving forward.”

Alon added, “We expect to see design happening in a more unified manner — generated by machine, validated by a person. We expect that to be the working model.” He also saw that it had to be extended from the physical domain into architecture and the system level.

Both agreed that the process has to start with a specification, which should be the golden reference of design intent. From that, a verification infrastructure should be created, but the human must stay in the loop. They also agreed that today only humans can do specification validation, and that decisions and assumptions are constantly verified by humans. AI is a productivity tool. “The IQ part of humans, the knowledge, the experiences that humans bring to the table continue to be significantly important,” asserted Abhi.

The resource shortage is being helped in the software world. “The existence of copilots isn’t going to eliminate all the software engineers in the world,” said Alon. “It is the other way around. Today, those copilots have 100 million users. In five years they will probably have 1 billion. People who don’t know how to code software will start coding software, because creating software is going to be an easier process. I hope that something similar will happen to the verification engineers. They will not have to worry about running the tools and debugging something. They can delegate that to somebody else.”

With certain tasks removed from an engineer’s daily routine, they can concentrate more on innovation. “I’m asking each one of my team members to start thinking about innovation,” Alon noted. “Start thinking about how they can harness not just the current EDA tools and the LLMs, but building a flow that we can use in the future. It’s more versatile, and there’s a lot of room for each and every engineer in the world to innovate. Engineers will be forced to become creators, problem solvers. They should be defining where they want to go and what they want to build, rather than how to build it.”

If the intent is to allow more innovation, can AI help with that? Real innovation means doing something that is different from what you have done in the past. Does that mean we might rely on AI hallucinations for innovation? The panelists talked about various programs that are looking into AI discovery and being able to harness hallucinations.

Abhi also talked about diffusion models. “Diffusion models purposely introduce noise in the data and reverse engineer that noise to see if it can correct itself and get it back to the original state. There are enough things that are being thought about to essentially address hallucination beyond the availability of high-quality data. Hallucinations are real. It is a risk, but there are people who are creatively using hallucinations. However, there needs to be some self-corrections that we are thinking about to detect these.”

Alon pointed out that hallucinations interfere with predictability. “I want excellence, and so there is no place for hallucination, but hallucination might help us get to new places. I have seen one example where a team was working on a new design for a radar. We asked AI to come up with the new architecture from scratch. It created an architecture the engineers could not understand, but it was working, and it was better — superior to their own.”

Abhi agreed. “We have to trust in that hallucination, that innovation to provide us with things that we’ve never seen before, and to have trust that our verification processes will eliminate the ones that were not viable hallucinations and actually promote the ones that were.”

As with all good panels, the audience was fired up with questions for both of them, and many of them concentrated on the practicalities of today: What tools were available, how were they being used, what were the problems that have been found. One interesting avenue explored the notion of interactivity with AI to help with design issues, especially as part of a high-level synthesis flow. It was thought that various agents would be involved, one that was perhaps aware of software, another about structure, and yet another about power. All of those agents would be working together throughout the design flow, and with each change and iteration the agents would be flagging any issues they find or making suggestions to the user. As we gain confidence in agentic flows, issues such as the readability of an intermediate description will become less of an issue than it was in the past.

There will be a transition period involving tools and the skill sets of people, as well as the infrastructure available within companies. Alon said this is moving very slowly in comparison to the advances in AI, but it’s improving. Others lamented about how important parts of an infrastructure are missing today — simple things like the ability to import PDFs and ask questions about them.

The panelists concluded that good progress is being made. Agents are working and they require a human to orchestrate them, and that relationship is likely to continue into the future. Companies have to be aggressive in the adoption of this technology, or they will get left behind and quickly find that the only things they can create are legacy, commoditized devices. “We all have to learn each and every day,” said Alon.

Related Reading
Iteration And Hallucination
For many aspects of an EDA flow, hallucinations from AI are not really that serious, because that is no worse than engineers on a Friday afternoon.