• Proteams Information Tech

AI is designing... AI chips?

The chips that are custom designed can take months of work to complete by humans, but the AI can produce these designs in under six hours


Yes, you read that title correctly, robots are helping build themselves. For years now, artificial intelligence has been worked on by Google, whom have been attempting to allow machines to learn how to create and explore themselves, but in a recent paper from Nature Machine Intelligence, a publication that is dedicated to machine learning and AI, created in response to the machine learning boom in the 2010s, have noted that they have been applying the research to a product for commercial use. Google’s newly announced TPU (tensor processing unit) chips that are optimized for AI, the previously designed artificial intelligence software are basically upgrading their own systems, quicker than humans and apparently more concisely.


The chips that are custom designed can take months of work to complete by humans, but the AI can produce these designs in under six hours and are easily comparable or even superior to those made by man in regards to power consumption, chip density and performance. There is a software tool called RePlAce that completes designs with speed akin to this new algorithm, but it failed to meet the standards on all tests compared with both humans and AI. In the paper, which was co-led by Azalia Mirhoseini and Anna Goldie, research scientists for Google, engineers noted that the work has “major implications” for the chip industry and will allow companies to customize chips for specific assignments.


Moore’s law stated that the number of transistors on a chip doubles every two years, this was predicted by Gordon Moore in the 1960s and ended up being a widely set target for the semiconductor industry. The end of this is forecasted sooner rather than later, with augmentations being made quicker than ever as to how to improve and downsize the necessary numbers. The newer AI software may not be able to help with the physical side of the transistors in the chips, but it will certainly aid in finding other ways to increase performance at the same rate.

Floorplanning is one of the stages of chip design, and this is what the AI has been tasked with. A time-consuming and very manual job for human designers previously, they would need to work with fiddly computer tools to find the most efficient layout on a silicon die for the chips’ sub-systems. The components are connected by miniscule wiring and will include things like memory cores and CPUs and the placing of them on a die even the slightest nanometer difference can end up having negative effects. Where each component is placed, it will affect the overall speed and efficiency of the chip.


The software started by developing solutions at random and were tested based off the performance and efficiency by a separate algorithm, which then fed back to the first. By doing it like this, it taught itself what strategies were optimal and built upon past experience.

“It started off kind of random and gets really bad placements, but after thousands of iterations it becomes extremely good and fast,” said Anna Goldie.

AI has been able to not only learn but outperform humans at games such as Go and chess, which is how they implemented the floorplanning task, they treated it like a game for the AI. The silicon die and it’s components are basically in place of game pieces, and the chip design “prize” for winning would be it’s overall efficiency. The engineers at Google trained a reinforcement learning algorithm on a dataset of 10,000 chip floorplans or different qualities, including randomly generated ones. The designs had specs for it’s reward based on power usage and length of wire required. The AI would then use the data to learn the good and bad floorplans and designed its own. The latest version of the TPU was used for the experiments and is designed to run the same type of neural network algorithms for the search engine and automatic translations. This means the new AI-designed chip will be used to design the next, and that one will design its own replacement, so on and so forth. The approach will be applied to other stages of chip design, which are also time-consuming, and could possibly decrease the overall design time down to just days.


“There’s a high opportunity cost in not releasing the next generation. Let’s say that the new one is much more power efficient. The level of the impact that can have on the carbon footprint of machine learning, given it’s deployed in all sorts of different data centres, is really valuable. Even one day earlier, it makes a big difference,” says Goldie.

There is a lot to be excited for in the future of artificial intelligence, but this could also mean humans could be out of jobs due to the increase.




Keep up-to-date with the latest tech industry insights, trends as well as information technologies, app development, and small business content with the Proteams Blog

Follow us on LinkedIn for updates on the latest tech news here

9 views0 comments

Recent Posts

See All