Google declared that it had developed AI software that can design computer chips faster than humans.
On Wednesday, the company’s researchers published a new paper in Nature. They stated that an edge graph convolutional neural network architecture learned how to create the physical layout. As a result, it allows artificial agents to perform chip design, and they are more experienced than any human designer.
Google used artificial intelligence to design other AI chips with more performance. This is meaningful progress in chip design. According to researchers, chip floorplanning has challenged automation despite five decades of research. It required months of serious effort by physical design engineers to create manufacturable designs. In contrast, they claim that their method automatically generates chip floorplans. They are superior or similar to those produced by humans in all principal metrics, including power consumption, performance, and chip area.
Google’s AI learning system is equipped with a learning program trained on a dataset of 10,000 different floorplans. They can create manufacturable layouts much faster.
Chip making for artificial intelligence is like a game
Systems like this can defeat humans at complex games such as chess and Go. This is because algorithms are trained to move pieces to increase their chance of winning strategically. As for the chip, the AI is trained to obtain the best combination of components to make it computationally effective.
The mechanics are simple. There is an integrated circuit to which they must place small components. The goal is to find the right fit. If artificial intelligence fails that combination, they try another until they get the correct one. The funny thing about AI is that it can perform unimaginable combinations, which in turn help the component work properly.
The AI system agent was pre-trained on a set of 10,000 chip floorplans to determine what works and what doesn’t. Each design has been labeled with a score. The algorithm used this data to generate the correct blueprints to build the chips.
Yann LeCun, Facebook’s chief AI scientist, praised the research on Twitter. As for the Nature editorial, the article hailed the breakthrough as a significant achievement.
While the results are not surprisingly different from human achievement, the speed at which artificial intelligence can make AI chips seems to be a fair reason to use this method.
Google has shared an image with the results obtained through AI. The stamp was blurred because the design is confidential. Nevertheless, it showed the difference between a chip designed by humans and a blueprint developed by artificial intelligence.