Chip design is an incredibly complex endeavor. So much so that only a handful of companies have found success in this arena. But what if designing a chip didn’t have to be so hard, and could artificial intelligence (AI) be part of the solution?

That’s the question Synopsys CEO Aart de Geus posed during his Hot Chips 2021 keynote speech this week.

“Every vertical market right now is investing AI to become a little smarter, more efficient, more effective,” he said. “Can AI design chips? The answer is yes.”

For those that aren't familiar with the company, Synopsys is responsible for developing some of the foundational software used to design semiconductors. Its platform allows some of the largest chipmakers in the world to design the nano-scale transistors required to develop smaller, faster, and more power-efficient chips by translating human-readable code into microscopic schematics.

And increasingly, this software is turning to machine learning and AI to do it.

Machine learning is nothing new for Synopsys. “Every one of our products at this point in time has some machine learning that amends the traditional algorithms,” de Geus said.

The algorithms are what allow the software to translate human-readable code into something that can be manufactured in any number of foundries around the globe. Synopsys has been using machine learning to boost communication speeds and reduce latency. However, this encompasses only a minutia of what’s possible, de Geus explained.

“But what can you do on the design flow? That is really the big question,” he said. “Is it possible to use AI, on a substantial number of design steps?”

And that’s exactly what Synopsys’ DSO.ai platform attempts to do. The software allows engineers and chip architects to explore possibilities beyond their own imaginations to find novel designs and combinations that eke out greater efficiencies or allow for higher clock speeds.

In his keynote, de Geus compared DSO.ai to a hunting dog tracking down an animal — the animal being the optimal combination of performance, power, and surface area. This by itself is already yielding significant gains in performance or voltage reductions, he claimed.

However, this is only part of the equation. By training the AI model, or in de Geus’ analog the hunting dog, it can lean on prior experiences to help it find the prey faster. In effect, with every new chip, the model gets a little smarter.

This, de Geus explains, translates into higher performance per watt and reduced surface area, which directly translates to higher yields. In some cases, he claimed the optimizations enabled by the software can result in a 26% performance improvement, which is the equivalent to a full-process node of progress.

What’s more, he claims the software can substantially reduce the time and talent required to bring leading-edge chips to market by allowing a single engineer to achieve what in the past would have taken an entire engineering team.

According to de Geus, this is only the beginning of what’s possible for the platform. Moving forward, the company is looking to expand the number of variables its AI design software can accommodate.

The software itself is limited by the number of variables the AI model has access to. The more variables the model has to work with, the more unique combinations it can test in its search for the ideal mix of power, size, and performance.

“The number of variables are enormous,” he said.

One area the company is already looking into is clock speed. de Geus described one trial in which it allowed the AI model to change the chip's clocks. This, he explained, allowed the software to find additional opportunities for optimization elsewhere rending a substantial improvement in power consumption.

However, clocks speeds are only one variable on Synopsys’ roadmap. Long term, the company aims to take software functions into account.

"Maybe at some point in time, it's even possible to revisit some aspects of functionality, maybe go as far as touch the software that's running on the chip," de Geus said.

In this scenario,  the DSO.ai would enable engineers to specify applications and workloads to optimize for.