The configuration showed performance comparable to solutions based on NVIDIA H100

The Rubytech group has completed testing of Russian large language models on Chinese graphics accelerators. Tests of Cotype 3 developed by MWS AI were conducted as part of the proprietary hardware and software complex "Machines of Artificial Intelligence Skala^r" with eight Chinese GPUs.

Engineers tested the model's operation without using NVIDIA accelerators, evaluating query processing speed, response latency, and system stability under increasing load. In the most common scenarios, the configuration showed performance comparable to solutions based on NVIDIA H100.

Additional optimization of drivers, software environment, and orchestration tools allowed for a 2–2.2 times increase in performance in certain scenarios compared to the initial environment. The containerized approach also simplified model deployment.

Rubytech believes that the results confirm the possibility of creating corporate AI systems on an alternative hardware base. This approach can reduce dependence on a single vendor and give companies more options when choosing equipment, considering cost, performance, and scaling prospects.

In the future, the company plans to test the platform with new generations of Chinese GPUs and expand the list of Russian AI services compatible with the system.