AI Can Design Microchips Better and Faster than Humans

Artificial intelligence (AI) can create computer microchips that operate at least as well as those developed by human professionals, and it can do it thousands of times quicker than humans.

Google‘s latest study is already assisting in developing microchips for the company’s next generation of AI computer systems.

Floorplanning, the process of creating the physical arrangement of a chip’s elements, is critical to a device’s eventual performance. This problematic endeavour frequently takes months of dedicated work from professionals, and after five decades of research, no automated floorplanning system has yet to achieve human-level performance.

Traditionally, designing microchips has been a complex and time-consuming task requiring extensive expertise in circuitry design. However, this new AI-powered system revolutionizes the process by streamlining it significantly. The researchers fed the AI system with terabytes of chip designs as training data, enabling it to learn intricate patterns and understand the underlying principles behind successful designs.

Using machine learning techniques, AI can sift through mountains of data in order to find trends and improve designs in real-time. This has the potential to improve material utilization, cut costs, and shorten the introduction of new products to the market. In addition to enhancing the quality and dependability of semiconductor designs, AI may also aid in identifying potential design defects or errors before they become severe concerns.

Because AI “learns from experience,” it outperforms other automated technologies.

According to study co-lead author Azalia Mirhoseini, a computer scientist at Google Research’s Brain Team in Mountain View, California, “the researchers had AI software interpret floorplanning as a game.

The chip’s processing cores, memory controllers, and other parts are the pieces of this technique. The canvas on which they sit on the board and the end outcome are optimal for various parameters such as chip size and power consumption.

The AI got quicker and better at chip creation as it studied a more significant number and diversity of microchips. This tactic outperformed earlier automated procedures “because it learns from experience,” according to research co-lead author Anna Goldie, a Google Brain Team computer scientist. “Previous methods failed to learn anything from each chip.”

According to the researchers, the new approach can design manufacturable chip floor designs in less than six hours, equivalent to or superior to those created by professionals in all critical parameters, including performance, energy consumption, and chip area. For each microprocessor generation, this may save hundreds of hours of human work.

“Even if a number of these designs were not intuitively superior to specialists,” Mirhoseini added, “they nonetheless generated pretty good outcomes.” Experts have come up with new ideas due to the AI-designed floor layouts, such as arranging components in doughnut shapes to cut distances between them, she said.

Google is currently producing these AI-designed circuits. According to the researchers, more powerful AI-designed technology might fuel additional AI advancements in future.