T1 IT maintenance tested more than 20 AI models on Chinese graphics accelerators, Kommersant writes. The cost of executing queries turned out to be 2-2.5 times lower than on Nvidia chips, with similar quality of answers, but in computer vision tasks, alternative solutions were inferior in speed. According to the company's calculations, replacing one partially charged Nvidia H100 card with a Chinese equivalent when it is placed in the cloud can save more than 2 million rubles a year.