Chinese chipmaker Iluvatar CoreX says its new Tiangai 300 GPU can match, and in some cases beat, Nvidia's H200 in real-world AI workloads. The claims come from a live demonstration for Chinese media, which saw text-to-image and text-to-video generation using frontier models such as Z.ai's GLM-5.3.
The live test and the company's numbers
In a text-to-image test, the Tiangai 300 completed every task at 0.101 images per second, against an 87.5% completion rate and 0.086 images per second for the H200. The Chinese chip, however, took 0.64 seconds longer to complete the task.
The demonstration was run by the company itself and shown to friendly media, a circumstance that weighs on how the results should be read.
The only independent test
The claim finds partial support in a third-party test by state-backed Shanghai INESA. According to INESA, the Tiangai 300, unveiled in July, delivered 1.05 times the H200's throughput under typical deployment conditions. Separate reporting from Wallstreetcn says the chip beat the H200 by 15% in mixture-of-experts (MoE) throughput and cut first-token latency by up to 30% in long-text processing.
For training, the GPU reportedly delivered 1.33 times the effective throughput of the H100 when fine-tuning Alibaba's Qwen3.8-32B. That comparison is against the H100, not the H200.
The Chinese market context
The H200 is the most advanced Nvidia AI GPU approved for export to China under current U.S. rules. Beijing has discouraged domestic firms from buying foreign chips, which has fuelled local rivals such as Biren, MetaX, Enflame and Moore Threads, alongside Iluvatar CoreX.
Nvidia's grip remains strong. The CUDA software stack is a decisive selling point for Chinese firms, and giants such as ByteDance, Alibaba and Tencent have preferred Nvidia chips over Huawei's Ascend 950PR thanks to better inference capabilities. Iluvatar CoreX, meanwhile, held just 0.3% of China's domestic general-purpose GPU market in 2024, a figure that puts the challenge in perspective.
The company, listed in Hong Kong earlier this year, outlined a roadmap in January aiming to overtake Nvidia's Rubin platform within two years.
No comments yet. Be the first!