Google is using machine learning from its AI to help it design its next generation of electronic chips. While human engineers take months to find the best possible arrangement of the different components of a chip, AI achieves better results in just 6 hours of work.
For many years now, Google has been putting a high point in integrating AI into the vast majority of its applications and services. At its Google I/O 2021 conference, the Mountain View firm, for example, presented an AI-powered dermatological support tool, which will allow users to identify their skin problems using a few photos.
This time around, Google researchers decided to use the machine learning of their AI to design their next electronic chips, and more precisely their next TPU (Tensor Processing Unit). As a reminder, this is an integrated circuit specially designed by Google to improve artificial intelligence systems by neural networks. In other words, Google is using AI to accelerate AI chip development.
As Google explains, finding the ideal design of a chip to combine performance and execution speed usually takes months for engineers. This task, called“floorplanning”, consists of finding the optimal layout of the subsystems of a chip. However, according to the researchers of the Californian giant, in just 6 hours of work the algorithms were able to develop more efficient chips than those designed by humans.
Instead of coins like pawns, you have components like CPUs and GPUs. Here, the objective is therefore to find the best possible arrangement, in order to obtain optimal calculation efficiency.
Next, the researchers trained the AIs via a data set consisting of 10,000 chip plans of varying quality, some of which were randomly generated. A specific”reward” function has been assigned to each design performed, based on its success in different areas, such as energy consumption for example.
The algorithm was then able to discern the most efficient plans from the worst, which in turn could generate its own plans. Google has already adopted this system and intends to use it to reduce production costs and produce more efficient chips.