A recent study by Google and published in Nature showed how artificial intelligence can be used to improve current semiconductor product design methods, which are the result of more than 60 years of efforts by scientists and engineers. The document describes a software agent that can design microcircuits by sequentially placing the macroblocks of which it is composed. The agent, which is a deep neural network, is trained using a paradigm called reinforcement learning, where positive changes are captured as possible solutions, and negative changes are discarded, allowing you to build a kind of decision tree that is optimized at each step.
AI is not yet applied at all stages of microchip development, but the situation is likely to change in the coming years. Now we are talking about the planning stage of the layout, which is actually one of the most painstaking. In fact, developers place macroblocks on a chip – pre-prepared sections of the circuit, the location of which relative to each other and to the rest of the microcircuit components is of paramount importance from the point of view of performance and efficiency. Humans can only place macroblocks by following a certain system that they can understand, whereas AI does not have this limitation. As a result, AI can create a more winning option. As an example, the illustration on the left shows a layout made by humans, and on the right, an AI.
In addition, the work of the designers is a painstaking and time-consuming process that can take weeks or months. The AI described in the study can create layouts that are better than those designed by humans in a matter of hours, saving enormous amounts of time.
AI has even demonstrated the ability to solve placement problems it has never encountered before. The researchers explain this by the fact that the system was trained on a very large base of ready-made structures. At the same time, the options created by the program after training turned out to be better than the original ones developed by people.