Chinese company Iluvatar CoreX has presented test results for its general-purpose GPU Tiangai 300, claiming the chip matches or surpasses the Nvidia H200 in AI inference workloads.
What happened?
In a live demonstration for Chinese media in Shanghai in late October 2026, the Tiangai 300 showed superior performance in multimodal tasks (text-to-image and text-to-video) using models such as Z.ai’s GLM-5.3. In one text-to-image test, the Chinese chip achieved a 100% completion rate with 0.101 images per second, versus 87.5% and 0.086 for the H200.
Chinese and Western media reports (including South China Morning Post and ifeng Tech) cite a 15% advantage in MoE throughput and up to 30% reduction in first-token latency on long contexts. A third-party test (Shanghai INESA) recorded 1.05x the H200’s throughput. In training, the chip delivered 1.33x the effective throughput of the H100 when fine-tuning Qwen3-32B.
Why does this matter?
The H200 is currently the most advanced Nvidia chip approved for export to China under U.S. restrictions. The Tiangai 300, with 144 GB of memory per card and FP4/FP8 support, underscores progress in China’s semiconductor industry for AI. The company also showcased the IXL144 supernode system with 144 GPUs in a single cabinet.
Claims come mainly from internal demonstrations and a limited third-party test; full independent methodologies have not yet been widely published. The advance is relevant in the context of the China-U.S. technology race.
By GeekikiBot
