Chinese Iluvatar Tiangai 300 accelerator outperforms Nvidia H200 in tests

Chinese tech company Iluvatar CoreX has publicly showcased its new Tiangai 300 accelerator, presenting test results in real-world AI scenarios. The device demonstrated performance levels that matched and even exceeded the Nvidia H200, marking a significant development for the domestic AI industry.

• Testing showed a throughput 5% higher than that of the Nvidia H200.
• The accelerator performs exceptionally well in complex multimodal AI tasks.
• Tests utilized large language models such as the GLM-5.3 from Zhipu AI.
• The breakthrough represents a move toward greater autonomy in Chinese chip manufacturing.

Technical Capabilities of Tiangai 300

During the Shanghai demonstration, the Tiangai 300 achieved a throughput 1.05 times that of the US-manufactured Nvidia H200. Given that the H200 remains one of the most powerful accelerators permitted for export to mainland China under current US trade restrictions, these findings indicate that local hardware is becoming a viable alternative for high-performance computing.

Multimodal AI and Performance

The presentation emphasized complex tasks, specifically the generation of video and images based on text prompts. By utilizing Zhipu AI's GLM-5.3 large language model, the Tiangai 300 proved that it can handle heavy computational workloads with high efficiency. The consistency shown during these tests is critical for developers and data centers looking to optimize their infrastructure.

Impact on the Industry

After entering the Hong Kong Stock Exchange in early 2026, Iluvatar CoreX has been scaling its operations. This achievement not only bolsters the company's position in the AI hardware market but also provides Chinese enterprises with a strategic alternative to western technology, reducing reliance on Nvidia in a constrained global market.