Researchers develop a technique that reduces energy consumption in AI by 95%

Researchers from the BitEnergy AI firm have taken an important step towards energy efficiency in the field of artificial intelligence. Their new method promises to reduce the energy consumption required to train and run AI models by 95%, without compromising the accuracy or performance of these systems.

As they point out in Decrypt, the key to this innovation lies in a new way of doing multiplications of floating point numbers, an essential operation in AI calculations. Traditionally, these multiplications consume large amounts of energy. However, researchers have managed to decompose these complex operations into simple integer additions, which significantly reduces the processing load and, therefore, energy consumption.

The floating point concept allows computers to handle very large or extremely small figures efficiently. But the greater the precision in the calculations, the greater the energy consumed. The proposed algorithm, called L-Mul, simplifies these processes and achieves the same level of precision at a much lower energy cost.

However, for this system to be implemented practically, it will be necessary to have specialized hardware that is currently not widely available. Even so, the creators of this method are optimistic and hope that in the future chips can be adapted to take advantage of this innovative technique.

This advance joins the efforts of various entities that are looking for more efficient ways to handle matrix multiplications, a key mathematical challenge in the development of artificial intelligence.