Iluvatar CoreX pits Tiangai 300 against Nvidia's H200 in Shanghai demo

Iluvatar CoreX publicly demonstrated its Tiangai 300 GPU alongside NVIDIA's H200 at a "2026 Media Open Day" event in Shanghai on October 28, positioning its chip against the US company's high-performance AI processor in several AI workloads.

In the demo, the Tiangai 300 produced its first image about 4 seconds after receiving an image-generation command. Iluvatar CoreX said the chip recorded higher throughput than the H200 in video generation.

In a text-to-image test, the Tiangai 300 achieved a 100% task completion rate at 0.101 images per second, according to the company, while the H200 achieved 87.5% at 0.086 images per second. The Tiangai 300 took 0.64 seconds longer than the H200 to complete the task.

Iluvatar CoreX also said the Tiangai 300 reached 15% higher throughput than the H200 on mixture-of-experts (MoE) workloads, and claims it cut first-token latency by up to 30% in long-text processing. When fine-tuning Alibaba's open-source Qwen3-32B model, the company says the Tiangai 300 achieved 1.33x the effective throughput of NVIDIA's H100. The demo drew on Z.ai's GLM-5.3 model, according to aitimes.com.

The Tiangai 300 was unveiled in July 2026. In a benchmark released that month by Shanghai INESA covering four major AI models and three usage scenarios, the chip recorded 1.05x higher per-GPU throughput than the H200 in a specific AI inference setup separating input processing from output generation. NVIDIA released the H200 in late 2023 and said in August 2026 that it had begun shipments of the processor to China.

Earlier this month, Iluvatar CoreX announced that the Tiangai series had secured a deployment case involving 10,000 GPUs.