FIT researchers used artificial intelligence to save energy in computers

Artificial intelligence is not omnipotent, of course, but it can make a lot of people's work easier and save their time. The research team at the Faculty of Information Technology that used evolutionary algorithms to automate the design of computer component knows this. The algorithms can be used in battery-powered devices where optimisation of ongoing operations is needed to reduce energy consumption.

Vojtěch Mrázek, Radek Hrbáček, Zdeněk Vašíček, and Lukáš Sekanina from the Evolvable Hardware research group have long been focusing on the application of evolutionary algorithms. They are currently focusing on their application in automated design of computer circuits. What does that mean? "For example, let's have a look at an adder. We wanted for it to have as low power input as possible, even if it meant an occasional mistake in the addition. Such occasional mistake is not really a problem for example in processing pictures where various values of pixels are added. A slight inaccuracy is hardly noticeable with the naked eye," said Lukáš Sekanina. Therefore, the experts tried to come up with a solution that would be the best compromise between error rate and power input. Not only have they managed that, but using artificial intelligence, they found an enormous amount of similar compromise solutions and created a library of approximate computer components.

"Based on requirements, e.g. required power consumption, our software can choose the most advantageous implementation from the point of view of error rate or performance from the library. It is possible to come up with five to ten options manually, but using the AI, we managed to come up with 430," Sekanina added. He also noted that the library is freely accessible at the project's website.

What are the practical applications of this software? The software can be put to use mostly in battery-powered devices where the battery endurance is emphasised, and the accuracy of calculations is less important. Playing videos on a mobile phone can serve as an example. If the battery is running low, the device will prioritise saving the battery over the video quality. Another way to use the algorithm is for smart glasses with intelligent object recognition. That requires image processing that is demanding both in terms of computing power and in energy consumption. Without the circuit solution with low power input, the operations would not be possible with the small battery in smart glasses.

The solution was successful at a prestigious conference

The foursome of experts presented the software at the spring DATE conference in Lausanne in Switzerland, one of the biggest and most prestigious global events in the field of automated of computer systems design, such as circuits, processors etc. Interestingly enough, the conference is primarily focused on traditional software used to design hardware, and yet the evaluation committee found the contribution of the four Brno researchers focusing on artificial intelligence the most interesting one. "We were different from the others - we used artificial intelligence and made it come up with a design of adders and multipliers with low power input." It is also not very common for all the professional reviewers of the submitted papers to award similar scores. "However, that was the case for us, and we got great evaluation from everyone," said Sekanina. Moreover, their contribution found success among the competition of 794 papers submitted from all over the world. The committee then chose ninety contributions in the category of interactive presentations, the BUT team's among them. It then became assessed as the best one in the category.