What it takes humans months, the artificial intelligence of Google does it in six hours. It is the claim that Google makes about its AI capable of designing machine learning chips “comparable or superior” to those made by humans. After years of experimentation, we will see the first commercial product in this regard shortly: Google’s next TPU chips have been designed by an AI.
What Google uses its AI to design chips optimized for AI it is no secret. However, now it seems that experiments have been stopped and applied to real products. They have also taken the opportunity to publish a study in Nature where they explain the development.
The big advantage what AI seems to contribute when designing chips is the speed. According to Google, the great time savings involved in using the algorithm to design instead of humans can have important implications for the industry. In principle, it should allow speeding up design iterations for upcoming chips, as well as rapidly designing chips for specific uses for which they are optimized.
Optimizing the space of a chip
Where AI seems to have an impact the most is in the planning the placement of elements on the chip. This process is essentially that of choosing where on the surface of the chip each element goes (CPU, GPU, memory …). It is essential, since it directly affects the speed and efficiency of the chip depending on how far each element is from others.
While for humans this is a problem of months of effort, an artificial intelligence takes it as a game. It interprets each element of the chip as a piece of a game and seeks to place it in the place where it offers the most efficiency, always taking into account all the other pieces and multiple other factors. After a few hours, it offers the position where the most computational efficiency is offered of the set of elements in the given limit.
To train the AI, Google says that gave you the data of 10,000 higher and lower quality chip designs. Each chip was labeled according to its quality and taking into account values such as the length of the wiring required or the energy use. In this way, the AI learned which designs are good and which are not, and then generate its own.
Via | CNBC
More information | Nature