Companies explore how AI could change chip design

AI is proving to be a useful tool in designing the very computer chips that power AI.

Tech companies are investing heavily in exploring how breakthroughs in AI might streamline the complex process of designing semiconductors and potentially stave off a long-predicted end to Moore’s law—the famous maxim that the number of transistors on a chip doubles every other year.

Researchers at Google’s AI-focused DeepMind arm recently discussed how they are using AI to “accelerate” chip design by fashioning circuits with a structure similar to a neural network and optimizing for speed, energy efficiency, and size with reinforcement learning, or shaping systems by rewarding desired outcomes. Researchers from DeepMind won a programming contest focused on designing smaller chips with this method, and the Wall Street Journal reported in July that the company is exploring how algorithms can help transform the next generation of its own chips.

“Despite five decades of research, chip floorplanning has defied automation, requiring months of intense effort by physical design engineers to produce manufacturable layouts,” DeepMind researchers wrote in a 2021 paper on the topic. “Our method…has the potential to save thousands of hours of human effort for each new generation.”

Google isn’t the only company that is invested in algorithms’ potential to transform the process of chip production. Deloitte predicted that “the world’s leading semiconductor companies” will spend $300 million on AI tools for chip design this year, with that number projected to grow by 20% each year through 2026.

In earlier years, the AI employed in the chip design process had been mostly relegated to pattern matching computer vision, according to Gartner VP Analyst Gaurav Gupta.

“[In that process], you look at the chip design pattern and you find certain clips, and you make some improvements on those clips from a manufacturing perspective, so the designs are more robust without giving up anything on the performance and power,” Gupta told Tech Brew.

More recently, chipmakers have begun to tap reinforcement learning algorithms trained on thousands of different chip floorplans and tasked with maximizing efficiency.

That’s the field that Google DeepMind has explored, as have semiconductor giants like Nvidia. Other companies, like Synopsis and Motivo.AI, have developed their own tools around AI for chip design, according to Gupta.

“Overall, the chip design process is very complex and requires a lot of engineering hours and investment tools,” Gupta said. “So I think AI can definitely help streamline that process, reduce your man hours, reduce the investment, the number of days.”

Some startups and researchers are now starting to explore how the next wave of generative AI that’s taken hold in the past several months could further change the process of chip design, though the concept is still only in its earliest stages, Gupta said. He said that while there have been statements in the press about generative AI being potentially being used for chip design, “there is very limited evidence that I have seen so far.”

One piece of experimentation was a contest held recently by the online chip design community Efabless for participants to design chips with generative-AI prompts. But Gupta said that “one of the biggest issues is the model on which the generative AI is trained; is it proprietary IP or not?”—the same hurdle the technology encounters in some product design fields.

“With reinforcement learning…people have been able to do some…placement, like placing the sub-blocks within the design, but not creating your own designs or codes,” Gupta said. “This is where generative AI can be helpful.”