Russian scientists have found a way to reduce the load on data centers, media writes

MOSCOW, October 13 - RIA Novosti. Russian scientists have developed the HARP algorithm to reduce the load on data centers, the Izvestia newspaper reported.

“MIPT specialists have developed the HARP algorithm, which allows you to compress language models with virtually no loss of accuracy. The technology can double the speed of AI systems, as well as reduce the need for computing power and memory in data centers. The algorithm was successfully tested on the Llama and Qwen language models. With the most severe compression, the error decreased by 15%,” the newspaper writes.

It is noted that HARP and neural networks already compressed with its help are available for free use.

The newspaper explains that large language models require significant computing resources, and their work is largely limited by memory bandwidth. To reduce the amount of data, scientists use quantization—representing the model's weights in fewer bits. The stronger the compression, the higher the risk of decreased accuracy. And the HARP algorithm allows you to reduce these losses.

In addition, HARP does not replace existing quantization methods, but works on top of them. The development was tested on models of different scales.

As NTI expert on new materials and technologies Evgeny Vishnevsky told Izvestia, the main advantage of this approach is that it is built into existing compression methods and improves their results, and developers do not have to rebuild familiar tools.