Google has advanced its research in machine learning to the point that it has created an artificial intelligence that’s capable of designing chips to be used for AI computation.
In shorter terms, it’s AI meant to beget AI. As published in the journal Nature, Google’s research team formulated a graph placement methodology by means of a neural network.
The challenge addressed by the AI is related to chip floorplanning, which is basically determining the best layout for a given integrated circuit. To train the AI, the research team approached floorplanning as if it were a game and to let the AI find the most optimal way to achieve the win conditions.
The result was an AI whose algorithm can yield chip designs that are comparable or even superior to human-made designs. Even more impressive is the fact that the AI can generate these designs in just six hours, whereas humans would need months to perform a similar effort.
Impressed with the result, this will be Google’s first time to apply its research into its next generation of tensor processing unit (TPU) chips for commercial purposes.