Skala^r

Rubytech Tests Russian LLM Cotype 3 on Chinese GPUs Matching NVIDIA H100 Performance

Rubytech has successfully tested the Russian large language model Cotype 3 developed by MWS AI on Chinese graphics accelerators instead of NVIDIA hardware. The trials took place on the Skala^r AI machine software-hardware complex and demonstrated stable operation with inference performance comparable to configurations using NVIDIA H100 GPUs. Eight Chinese GPUs were used in the test system, delivering an average first-token generation time of about 8 seconds and inter-token latency between 111 ms when processing 27,000-token contexts. Engineers adapted drivers, software environments, and orchestration tools, achieving up to 2-2.2 times faster performance in certain scenarios compared to the standard Chinese GPU setup. The results indicate that corporate AI infrastructure can be built without mandatory reliance on NVIDIA accelerators, which is particularly relevant for government agencies, state-owned companies, and critical information infrastructure organizations requiring localized model deployment and controlled scaling. Rubytech also noted potential reductions in total cost of ownership with Chinese GPUs, although the company emphasized this is an additional option rather than a complete replacement for NVIDIA solutions.