For decades, constant innovation in the world of semiconductor chip design has made processors faster, more efficient, and easier to produce. Artificial intelligence (A.I.) is leading the next wave of innovation, trimming the chip design process from years to months by making it fully autonomous.
Google, Nvidia, and others have showcased specialized chips designed by A.I., and electronic design automation (EDA) companies have already leveraged A.I. to speed up chip design. Software company Synopsys has a broader vision: Chips designed by A.I. from start to finish.
A.I. has already shown its staying power in the world of semiconductors, and moving forward, it could be involved like never before. Here’s how.
Synopsys announced at Hot Chips, an annual semiconductor conference, an expansion of its DSO.ai software that can handle the entire chip design process. DSO.ai is already used by company’s like Samsung to design its Exynos chips for smartphones and other other smart devices. However, it currently only handles a narrow set of design challenges. Now, Synopsys says it can handle the process from start to finish.
Synopsys calls it “software-designed hardware,” which is a break from the “software-defined hardware” that chip designers have used for years. “You have tools where you take a chip from specification all the way to the final implementation that you ship to a foundry,” Stelios Diamantidis, senior director of A.I. solutions at Synopsys, said.
This process isn’t new, just expanded. DSO.ai already works on the layout portion of the design process, using A.I. to determine the optimal layout of components for power and performance. Google recently unveiled something similar, where it used A.I. to design one of its tensor processing units (TPUs). Basically, Google used A.I. to help design a chip that handles A.I. Nvidia has also invested in machine learning to help design chip layouts.
This advancement alone speeds up the design process, shaving down what would normally take many weeks to a matter of days — and generally with a performance advantage. Synopsys took that idea and expanded it. “Can we really design chips fast enough and cheap enough to become personalized?” Diamantidis said.
Time and money are the barriers into the world of semiconductor design, and they’ve been hurdles for new companies interested in the space. This is not only important for chips inside processors, graphics cards, and other PC components, but also the millions more that are necessary for smart devices, medical equipment, and cars — to name a few.
Synopsys uses a model called reinforcement learning, which trains the machine on how to achieve the best reward given a task. This model doesn’t identify patterns in datasets or predict outcomes. Rather, it finds the optimal path given large amounts of data through a reward system.
This is the same model Google used to train A.I. to win in Chess and Go, just infinitely more complex. “You’re looking at an unfathomable number of states, way more than the atoms in the known universe,” Diamantidis said. The DSO.ai solution only addresses one point of the design process, and it’s already “trillions of times more complex” than training A.I. in Chess or Go.
In terms of what A.I. can do for chip design, saving time is the most important. Synopsys says what would have normally taken two years in the past can now be done in as little as three to six months. It claims that this will help companies build chips faster, cheaper, and for more specialized purposes.
Synopsys has a portfolio of software that “literally everybody in the world uses to design chips.” Although that claim might be hyperbole, it’s true that Synopsys touches most parts of the semiconductor industry. Its website says Synospsys software is behind 90% of the world’s FinFET chip designs, and that’s only a small portion of what the company does.
In addition to being faster and cheaper, Synopsys claims that A.I.-designed chips are more efficient. In a press briefing, the company said it saw up to a 26% power savings using the A.I. model compared to a human engineer, which is a larger gain than moving to a new manufacturing process. “When people go from a manufacturing process to the next manufacturing process, let’s say from 7nm design to 5nm design, the scaling factor they’re looking for (is 20% at best),” Diamantidis said
In practical terms, says Karl Freund, industry analyst and founder of Cambrian A.I. Research, this will lead to faster innovation. “Over the next few years, we will consequently enjoy a faster pace of innovation, and better products with longer battery life,” Freund wrote in an email to Digital Trends.
Although Synopsys is leading the charge, it isn’t the only company looking toward A.I. for future chip design. Cadence, a competitor of Synopsys, has its Cerebrus tool that helps engineers automate multiple points in the chip design process. However, Synopsys and industry analysts say this tool is still lagging behind. “Synopsys has at least an 18 month lead in this area, perhaps more,” Freund of Cambrian A.I. Research said.
Larger companies, such as Google and Nvidia, have also invested in this space, though they’re not able to deliver a tool that handles the chip design from start to finish, according to Synopsys. “You can’t really build an autonomous car unless you build a car first,” Thomas Andersen, vice president of A.I. and machine learning at Synopsys, said.
We’re still in the early stages of A.I.-designed chips. Diamantidis thinks adoption will come in a number of waves, the first of which will help companies optimize their existing design workflows. By the third wave, Synopsys hopes to open “the door to, really, the folks who are not experts in chip design.”
Over the past few years, there has been buzz around the idea of Moore’s Law, where chips get faster and more efficient on a cycle of around two years. A.I. looks to be leading the charge into a new era of Moore’s Law, helping more companies build faster and more efficient chips, and at speeds that would otherwise be impossible.
For chips designed fully by A.I., it’s a matter of when, not if. Many of the world’s largest semiconductor companies are already using A.I. at various points in the design process, and some of the world’s largest tech companies have invested heavily in the technology. With a tool that can handle design from start to finish, chip design might be pushed over the edge into a revolution.
- Nvidia’s next GPUs will be designed partially by AI
- As Google keeps racing ahead, where is Apple’s A.I. strategy?