Google AI pushes itself into the new frontier of artificial intelligence, after machine learning, the ability to create and elaborate an intelligent design strategy, applied to one of the most difficult fields of engineering, the design of a chip.
Imagine that normally under optimal conditions, a team of IT engineers would take at least 6 months to design a chip, but from the Google AI paper published in Nature, the AI took only 6 hours for the same result.
Google AI: The AI-based method took a few hours to design the chip
Through an intense preparation work of machine learning, in which all the parts of the chip must be organized, we are talking about km of small cables, memories, CPU and GPU, optimized to fit within a limited available space, Google AI engineers have “taught” logistic methods to artificial intelligence of optimal organization, called “floorplanning”.
It must be said that this floor planning method is a kind of Pandora’s box, there is an understandably limited number of combination possibilities, we are talking about billions of possible outcomes. The method that engineers normally use, tries to reduce the possible possibilities to the most useful ones, but it is not always certain that the final solution is really the best ever, and still requires months of work.We start from macro blocks, i.e. the main components, and gradually we try to arrange all the logic gates in the available space, and if one of these is changed in position, just as many new possibilities will become accessible.The computational capacity of artificial intelligence designed by Google AI allows you to develop a superior strategy, that is, to discard all possible combinations, even if optimal, until you find the one, if you can say “perfect”, reducing the consumption of resources. to a minimum and a significant saving of funds, effectively making costs more accessible to the consumer. “In less than six hours, our method automatically generated chip floorplans that are equivalent to or better than people made by all major parameters, including energy efficiency, performance, and chip area.“It reads. Google has already used the AI method to configure a TPU, for the application of machine learning, loading 10000 floorplan presets already used and produced.