The chip was designed by an Artificial Intelligence (AI) "trained" by Google to reduce design time. The same Californian company has kept to specify that normally, such a project requires "months of intense work" . The results of the AI work were described in an article published in Nature .
The stage of work where AI has been most useful is the one in which, normally, the engineers deal with organizing all the components of the chip in a rather limited available space, the so-called floorplanning, with the relative connecting cables.
The results achieved
Floorplanning represents a phase that does not require a single solution, if anything an optimal one. The possible combinations in the positioning of the different elements are in fact billions and normally the engineers follow a precise positioning algorithm that starts from the “macro” components, those that take up more space, and ends with the standard ones. Clearly, the change in position of a single element causes the birth of a new possible configuration.
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 power efficiency, performance, and chip area.
How Artificial Intelligence works for the new chips
The AI exploits the concepts of machine learning to learn the different possibilities of floor planning that are used precisely to "train" artificial intelligence, which, following the creation of this database, is able to identify the different possibilities that allow to obtain the best configuration not only in terms of space occupied, but also in terms of energy dissipation and chip operation.
The idea was born after observing how artificial intelligences were able to compete, and beat, man in board games that involve a certain pattern, such as chess and Go. Instead of a board, the robot has to deal with a silicone nut. Instead of chess it has to work with components like CPU and GPU. The goal , therefore, is simply to find the "victory conditions" of each card, ie the computational efficiency.
This way algorithm designs are "comparable or superior" to those created by humans , Google engineers say, but they can be generated much, much faster. The use of AI will allow to accelerate the development of artificial intelligences themselves.
The effect of the use of artificial intelligence on the chip industry
An editorial also published by Nature defines the result as important not so much for itself, but for future applications for the chip industry .
However, it states that " technical expertise must be widely shared to make sure that the 'ecosystem' of companies becomes truly global ". He went on to emphasize that "the industry must make sure that time-saving techniques do not alienate people with the necessary core skills."
Our method was used in production to design the next generation of Google TPU
The use of this design methodology is destined to expand like wildfire, not only for Google, but also for all the other houses that deal with producing chips.
Such work could help offset the anticipated demise of Moore's Law, a 1970s chip design axiom that states that the number of transistors on a chip doubles every two years or so.
The complexity of a microcircuit, measured for example by the number of transistors per chip, doubles every 18 months (and therefore quadruples every 3 years).
AI won't necessarily solve the physical challenges of placing more and more transistors on chips, but it could help find other pathways to boost performance at the same rate. For now, the method has been used to configure a TPU from Google itself, that is, the processing units that deal with applications related to machine learning. But the use of AI in this field represents a sensational turning point from an economic point of view and development times.
The article Google relies on artificial intelligence: here are its new chips comes from Tech CuE | Close-up Engineering .