Artificial intelligence is now helps design computer chips – including those needed to run the most powerful AI code.
Drawing a computer chip is both complex and intricate, requiring designers to arrange billions of components on a surface smaller than a nail. Decisions at each stage can affect the final performance and reliability of the chip, so the best chip designers rely on years of experience and hard-earned knowledge to lay out circuits that push the best performance and energy efficiency from nanoscopic devices. Previous efforts to automate chip design over decades have been limited.
But recent advances in AI have made it possible for algorithms to learn some of the dark art involved in chip design. This will help companies prepare more powerful and more efficient drawings in a much shorter time. It is important that the approach can also help engineers design AI software, experiment with different code adjustments along with different circuit layouts to find the optimal configuration of both.
At the same time, the emergence of AI has aroused new interest in all kinds of new chip designs. Advanced chips are becoming increasingly important for just about every corner of the economy, from cars to medical equipment to scientific research.
Chipmakers, including Nvidia, Google and IBM, test all AI tools that help arrange components and wires on complex chips. The approach can shake the chip industry, but it can also introduce new technical complexities, because the type of algorithms that are distributed can sometimes behave in unpredictable ways.
At Nvidia, lead researcher Haoxing “Mark” Ren is testing how an AI concept known as reinforcement learning can help arrange components on a chip and how to connect them. The approach, which allows a machine to learn from experience and experimentation, has been the key to some major advances in AI.
The AI tools Clean try to explore different chip designs in simulation, and train a large artificial neural network to recognize which decisions ultimately produce a high-performance chip. Ren says that the approach should reduce the technical effort required to produce a chip by half while producing a chip that matches or exceeds the performance of a man-made.
“You can design chips more efficiently,” says Ren. “It also gives you the opportunity to explore more design space, which means you can make better chips.”
Nvidia began making graphics cards for gamers, but quickly saw the potential with the same chips to run powerful machine learning algorithms, and is now a leading manufacturer of advanced AI chips. Ren says that Nvidia plans to bring chips to the market made with AI, but declined to say how soon. In the distant future, he says, “you’ll probably see a lot of the chips designed with AI.”
Reinforcement learning was most commonly used to train computers to play complex games, including the board game Go, with superhuman skill, without any explicit instruction about the game’s rules or principles of good play. It shows promise for various practical applications, including training robots to understand new objects, flying fighter jets and algorithmic stock trading.
Song Han, an assistant professor of electrical engineering and computer science at MIT, says reinforcement learning shows significant potential to improve the design of chips, because, as with a game like Go, it can be difficult to predict good decisions without many years of experience and practice. .
His research team recently developed a tool that uses gain learning to identify the optimal size of different transistors on a computer chip, by exploring different chip designs in simulation. It is important that it can also transfer what it has learned from one type of chip to another, which promises to lower the cost of automating the process. In experiments, the AI tool produced circuit designs that were 2.3 times more energy efficient while generating a fifth as much interference as those designed by human engineers. MIT researchers are working with AI algorithms at the same time as new chip designs to get the most out of both.
Other industry players – especially those who are heavily invested in developing and using AI – also want to use AI as a tool for chip design.