20-year-old idea increases the efficiency of AI by 1,000 times

It is primarily data transfer that causes a large proportion of energy consumption. This is even more the case with the enormous amounts of data that artificial intelligence algorithms work with.

Omitting this step, which is also considered a bottleneck between memory and logical processes, therefore has an enormous impact. And this is exactly where an idea from 2003 comes into play, which was developed at the University of Minnesota back then.

This led to a collaboration between numerous disciplines from physics to engineering and computer science. The result was a number of circuits that are used today in smart watches and memory elements.

Computational Random Access Memory, or CRAM, was also involved. This allows the actual calculation and execution of parallel processes directly in the main memory and therein at any location.

More than just a new architecture

Furthermore, these are not conventional circuits, but magnetic tunnel contacts that can use the electron spin instead of the charge to switch between 0 and 1.

In applications based on artificial intelligence, this results in a power consumption of one thousandth in order to ultimately achieve the same result as with the classic method. The current and forecast electricity consumption of neural networks worldwide shows just how huge this amount of electricity is. According to the International Energy Agency, 460 terawatt hours were used in 2022. By 2026 at the latest, it is expected to be 1,000 terawatt hours.

With the gigantic savings potential of 99.9 percent, this would leave 999 terawatt hours that would no longer be needed. This corresponds to the annual electricity consumption of Japan, the fourth-largest economy with 126 million inhabitants.

According to the paper, this would not even be the best possible result. Further tests could reduce energy consumption by a factor of 1,700 or 2,500. This additional increase in efficiency is made possible by a targeted adaptation of the CRAM to individual algorithms, which can be calculated even faster and therefore more economically.

The above is the detailed content of 20-year-old idea increases the efficiency of AI by 1,000 times. For more information, please follow other related articles on the PHP Chinese website!