The study, published in the journal Nature, shows that it can take humans ‘months’ to design specialized chips for tensor processing units – the type of chip used in… Artificial intelligence But the Reinforcement Learning (RL) algorithm is better and faster than humans.
The researchers wrote in the study: ‘The RL agent becomes better and faster at optimizing floor planning because it sets up a greater number of network lists.
It appears that it can generate chip floor plans comparable to or superior to human experts in less than six hours, while it takes humans months to produce acceptable blueprints for modern accelerators.”
Google researchers gave the program 10,000 floor plans to slice it for analysis, and then figured out how to come up with floor plans that don’t use more space, wires, and electrical power than those designed by humans.
The floor plan of the chip is defined as where parts such as CPUs, GPUs, and memory are placed on the silicon.
And since the 1960s, there have been three different ways of how to put these parts on silicon: segmentation-based methods, stochastic approaches, and analytics.
None of them have achieved the level of human performance, but the RL system is able to do so fairly easily.
“Our method, on the other hand, can expand to lists of networks containing millions of nodes, and directly optimize for any combination of differentiated or non-differentiable cost functions,” the researchers added.
In addition to the immediate impact on chip floor planning, the ability of our method to generalize and quickly generate high-quality solutions has significant implications, opening opportunities for co-improvement with earlier stages of the chip design process.
The feat is widely regarded as remarkable, and has been lauded by some of the world’s leading AI researchers, including for being from Facebook.