Several researchers from the University of California, Berkeley, and Google AI claim to have found a way to harness artificial intelligence (AI) to design smaller and smaller chips faster. In one blog post published by the Google subsidiary, the experts said they had developed a deep learning algorithm to achieve this.
A step towards the use of artificial intelligence for the semiconductor industry
While the world is in the midst of a shortage of semiconductors, several scientists are trying to reduce the manufacturing time of electronic chips. The process mostly practiced by semiconductor giants involves up to 1,400 steps. It sometimes takes up to 20 weeks to manufacture the most efficient chips, engraved in 5 or 7 nanometers, i.e. 10-9 Mr.
Amazon launches AWS for Games for game developers
Last year, IBM took an additional step by successfully developing a chip engraved on 2nm. With a desire to reduce the size of chips, while making them more efficient, companies are forced to develop new, ever more sophisticated manufacturing techniques.
With its area of expertise, Google AI has already thought about the design of electronic components via AI. Last year, Mountain View claimed it could develop chips in just six hours using machine learning. This process aims to train a model so that it gains in maturity and can correctly reproduce a process down to the nanometer.
A few months later, Google subsidiary researchers Amir Yazdanbakhsh and Aviral Kumar continued to work on the use of AI in chip manufacturing processes. They designed the PRIME model, based on deep learning.
Deep learning at the center of the PRIME model developed by Google AI
When designing electronic components, it is necessary to use patterns. They help to avoid errors during the chip creation process. The PRIME model makes it possible, thanks to the data offered to it, to generate architectures of electronic chips without having to use these patterns. A time saver, used to reuse data used in a previous manufacturing session instantly.
Thanks to the PRIME approach, it is possible to design chips with up to 50% lower latency than traditionally manufactured chips. Image: Google AI.
Based on the data provided by researchers, PRIME takes into account both the correctly manufactured chips, those with the best performing characteristics, but also the imperfections, to avoid making the same mistakes. The data that led to the design of a high-performance chip are remobilized by the model to offer a chip that is slightly more efficient, or at least just as efficient.