Iluvatar CoreX says its Tiangai 300 GPU matches Nvidia's H200 on AI inference

Iluvatar CoreX, a Shanghai-based developer of general-purpose GPUs, used a media open day in Shanghai in late October to show its Tiangai 300 chip running live next to Nvidia's H200. The company says the GPU matches, and in some tasks slightly exceeds, the Nvidia part on real AI inference workloads. The figures come from the company and a single outside test whose method has not been published.

What the company showed

The headline number comes from a test by Shanghai INESA, a state-backed firm, which put the Tiangai 300's normalised per-card throughput at 1.05 times that of the H200, according to TokenHeart and Wccftech. TokenHeart notes that this metric averages a multi-card result per card and says nothing about latency or peak performance, and that Iluvatar has not disclosed the workloads or test conditions. City News Service reports the result was obtained under standard disaggregated deployment settings.

In the live demonstrations, the chip ran text-to-image and text-to-video tasks with large models including Zhipu AI's GLM-5.3. In one image test cited by Wccftech, it completed all tasks at 0.101 images per second, while the H200 finished 87.5% of them at 0.086 images per second.

Iluvatar also presented the IXL144, a rack-scale "super node" with 144 Tiangai 300 GPUs, 18 CPUs and about 10.1 TB of memory. By the company's own calculation, it lowers the cost per million effective tokens by 40% compared with unnamed mainstream international alternatives.

Context and caveats

The Tiangai 300 was first announced in July at the World Artificial Intelligence Conference, where Iluvatar compared it with Nvidia's Hopper generation rather than naming the H200. The company began trading in Hong Kong in January, according to City News Service, and Wccftech cites a 0.3% share of China's general-purpose GPU market in 2024.

The H200 is the most capable Nvidia AI GPU approved for export to China under current US controls, which makes it the natural benchmark for domestic rivals such as Biren, MetaX and Moore Threads. Wccftech notes that Beijing has discouraged purchases of foreign chips, while many Chinese firms still rely on Nvidia hardware partly because of its CUDA software ecosystem. No independent lab has yet published a reproducible comparison, so the parity claim should be treated as unverified until fuller results or customer deployments emerge.