Can artificial intelligence do a better job designing microchips than human specialists? The Google Brain Team tried to answer this question and came up with some interesting conclusions. It turns out that a well-trained artificial intelligence can design computer chips – and with excellent results. So excellent, in fact, that the next generation of A.I. computer systems from Google will include microchips created using this experiment.
Azalia Mirhoseini, one of the computer scientists on Google Research’s Brain Team, along with several colleagues, explained this approach in an issue of Nature. Artificial intelligence usually beats human intelligence easily when it comes to games such as chess. Some might say that artificial intelligence can’t think like a human, but in the case of microarrays, this has proven to be the key to finding out-of-the-box solutions.
Toluna launches industry’s ‘first-of-its-kind’ enterprise insights subscription in 2021
Designing a microchip involves “step-by-step planning,” a lengthy process that involves human experts using computer tools. The goal is to find the optimal arrangement of all subsystems on the chip, thereby ensuring the best performance. Small changes in the placement of each component can have a huge impact on how powerful the chip will be, whether it’s the processor, the graphics card or the memory core.
Google engineers admit that developing a plan for a new microchip takes “months of hard work” by an entire team of people. But Google Research’s Brain Team, based in Mountain View, California, seems to have cracked the code that simplifies the whole process. The answer? Treating floorplanning like a game.
According to Azalia Mirhoseini and Anna Goldy, co-directors of the research team, the AI has been trained to play a game of finding the most efficient chip design. Using a dataset of 10,000 microchip plans, the team applied a reinforcement learning algorithm to separate good plans from bad ones. Metrics such as wire length, power consumption, chip size and others were taken into account.
The more the AI was able to identify the most optimal chip configurations, the more it was able to create its own. In the process, it found some unique approaches to part placement. This inspired experts to try something new, such as reducing the distance between components by arranging them in a donut shape.
Although earlier attempts had been made to simplify the process, five decades of research had yielded no solutions. So far, all automated scheduling methods have been unable to replicate the performance that human-made chips have provided.
According to Anne Goldy, this is because the algorithm learns from experience. “Previous approaches didn’t learn anything with every chip,” Goldy says, pointing to the use of machine learning.
“What used to take a team of experts months, can now be replicated by artificial intelligence in less than six hours. The resulting microchip plans are as good as human ones, and in some cases superior.” Thus, Google’s new discoveries could save hundreds, if not thousands, of working hours for each new generation of computer chips.
The company is currently using these A.I. chips for further research. Scientists speculate that using these more powerful chips could further advance research, including using A.I. for such things as vaccine testing or city planning. As A.I. becomes more widespread, more major discoveries are sure to come our way in the near future.