SHANGHAI — Chinese artificial intelligence chipmaker Iluvatar CoreX has demonstrated its new Tiangai 300 graphics processing unit (GPU), claiming that the processor matches and, in some tests, exceeds the performance of Nvidia's H200 in selected AI workloads.
The company demonstrated the chip during a media open day in Shanghai on October 28, 2026. The H200 is currently the most advanced Nvidia AI processor approved for export to mainland China under U.S. trade restrictions.
Iluvatar CoreX said the Tiangai 300 is the first domestic general-purpose GPU to achieve performance parity with the H200 in third-party testing.
Tiangai 300 Performance Compared with Nvidia H200
A benchmark conducted by Shanghai INESA, a state-backed organization, measured the Tiangai 300 at 1.05 times the H200's per-card end-to-end throughput under standard deployment conditions. The tests used prefill-decode disaggregated configurations commonly applied in large language model inference.
During the live demonstration, the Tiangai 300 was tested on multimodal AI tasks, including text-to-image and text-to-video generation using models such as Z.ai's GLM-5.3.
In a text-to-image test, the Tiangai 300 achieved a 100% completion rate and throughput of 0.101 images per second. The Nvidia H200 recorded an 87.5% success rate and processed 0.086 images per second. However, the Tiangai 300 took 0.64 seconds longer to complete the task in that comparison.
Additional results reported by the company include:
-
15% higher throughput than the H200 for mixture-of-experts workloads using GLM-5.3 at a 64K context length.
-
10% to 30% lower first-token response latency in long-text processing tasks with context lengths ranging from 8K to 128K.
-
Attention performance at 64K context reaching 1.05 times the H200.
-
Multimodal generation throughput approximately 2% to 5% higher in some tests.
-
21% higher recommendation inference throughput in online services such as search, advertising and recommendation.
In an eight-card FP8 training demonstration, the Tiangai 300 delivered 1.33 times the effective throughput of Nvidia's H100 while fine-tuning Alibaba's open-source Qwen3-32B model.
Tiangai 300 Specifications and IXL144 Supernode
The Tiangai 300 features 144 GB of memory per card, compared with 141 GB on the Nvidia H200. It supports lower-precision computing formats, including FP4 and FP8. The chip was first introduced in July 2026 at the World Artificial Intelligence Conference.
Iluvatar CoreX also presented its IXL144 supernode system, which integrates 144 Tiangai 300 GPUs and 18 CPUs in a single rack, with approximately 10.1 TB of memory. The company estimates that the configuration can reduce the cost per one million effective tokens by around 40% compared with mainstream international solutions.
The company has also said its Tiangai series has reached deployments involving 10,000 cards.
China's Push for Domestic AI Chips
Iluvatar CoreX is among several Chinese companies developing domestic alternatives to advanced U.S. AI processors as export restrictions limit China's access to high-performance computing hardware.
The company began listing in Hong Kong in 2026. Its Tiangai 300 results indicate progress in China's development of domestic AI processors for large-scale commercial applications.
However, the reported performance figures should be considered in the context of the specific tests. Full details of the benchmark methodologies, model configurations and workload conditions have not been publicly disclosed, and the results have not been independently reproduced across a broad range of conditions.
——— End of Article ———