Google claims to have developed artificial intelligence software that can design computer chips faster than humans.
Google researchers have published a new article in Nature on Wednesday describing “a graph-based convolutional neural network architecture” that learned how to design the physical layout of a semiconductor in a way that allows “the chip design to be performed by artificial agents with more experience than any human designer “.
The researchers describe a deep reinforcement learning system that can create floor plans in less than six hours, while human engineers and their automated tools can take months to arrive at an optimal layout.
See how the researchers described their achievements in the article summary:
Despite five decades of research, chip planning has defied automation, requiring months of intense effort from physical design engineers to produce manufacturable layouts. Here, we present a deep reinforcement learning approach to chip planning. In less than six hours, our method automatically generates chip floor plans that are superior or comparable to those produced by humans on all key metrics, including power consumption, performance and chip area.
In terms of design, Google is referring to crafting a chip’s floor plan, which is the arrangement of its subsystems – like its CPU and GPU cores, cache memory, RAM controllers, and so on – into its mold of silicon.
Google has already used this AI system to produce the floor plan of a next-generation TPU – its Tensor Processing Unit, which the web giant uses to speed up neural networks in its search engine, public cloud, AlphaGo and AlphaZero and other projects and products, notes the report.
In fact, Google is using machine learning software to optimize future chips that accelerate machine learning software. The virtuous cycle of AI to AI design chips looks like it’s just getting started.