T1 tests more than 20 SI models on Chinese accelerators
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Matching an accelerator to a specific task can substantially reduce the cost of running SI, although platforms differ in their advantages for speed and fine-tuning large models.
According to the company, its laboratory conducted more than 1,000 runs using GPUs from different manufacturers. In document-processing and knowledge-base search tasks, Chinese accelerators reduced inference costs by 2–2.5 times while maintaining comparable answer quality. For fine-tuning large models, T1 proposes distributing the workload: retaining Nvidia for training and the largest models while moving suitable services to alternative chips. In computer vision, Chinese solutions were price-competitive but slower than Nvidia. The company also estimated that replacing an underutilised H100 with a single C550 card in the cloud could save more than 2 million rubles per year; this is the result of its own tests and a specific configuration.
Sources 1