China’s Tiangai 300 GPU Claims Edge Over NVIDIA H200 In Internal Demo

Iluvatar CoreX, a Chinese chipmaker, claims its Tiangai 300 GPU can match or surpass NVIDIA’s H200 in AI workloads, based on a live demonstration for Chinese media. The test involved text-to-image and text-to-video generation using models like Z.ai’s GLM-5.3.

In one text-to-image test, the Tiangai 300 completed all tasks at 0.101 images per second, compared to an 87.5% completion rate and 0.086 images per second for the H200. However, the Tiangai 300 took 0.64 seconds longer to finish the task.

A third-party test by state-backed Shanghai INESA found the Tiangai 300 delivered 1.05 times the throughput of the H200 under typical conditions. Reports from Wallstreetcn noted a 15% throughput advantage in mixture-of-experts (MoE) tasks and a 30% reduction in first-token latency for long-text processing.

In training, the Tiangai 300 reportedly provided 1.33 times the effective throughput of the older H100 when fine-tuning Alibaba’s Qwen3-32B. The H200 is the latest NVIDIA GPU approved for export to China under U.S. rules. Domestic firms are also being encouraged to avoid foreign chips, leading to competition from companies like Biren and Moore Threads.

Despite these efforts, Chinese companies still use NVIDIA GPUs, partly due to the CUDA software stack. A security firm found 32 B300 units in China Telecom, likely smuggled in. Major firms like Alibaba and Tencent continue to favor NVIDIA’s RTX PRO 5500 over Huawei’s Ascend 950PR for inference capabilities.

The benchmark results largely come from a company-run demo for friendly media. The only external test, from INESA, shows a 5% edge for the Tiangai 300, which is not a significant lead. Iluvatar CoreX holds just 0.3% of China’s domestic general-purpose GPU market. The firm, however, has listed in Hong Kong and outlined a plan to surpass NVIDIA’s Rubin platform within two years.