At its October 28 'From Model to World' media open day, Tianshu Zhixin (Iluvatar CoreX) debuted live demonstrations of its Tiangai 300 AI chip, outperforming Nvidia's H200 in GLM-5.3 large-model, text-to-image and text-to-video scenarios, Fast Technology reported on October 29.
Based on the company's Tianshu architecture with a new-generation Tensor Core, 144GB of memory and native support for FP4, FP8, FP16 and BF16, the chip showed an 8-card FP8 training demo on Qwen3-32B with per-card throughput 1.33 times that of H100.
Third-party testing by Shanghai Electric completed in July found that under typical PD-separated deployment the Tiangai 300's single-card end-to-end throughput reached 1.05 times H200's - the first domestic general-purpose GPU measured at H200 level - with agent first-token latency 10% to 30% lower across 8K to 128K contexts.
The company's IXL144 supernode packs 144 GPUs in a standard cabinet with about 10.1TB of memory, and its per-million-effective-token cost is roughly 40% lower than mainstream international solutions, marking a step from 'can it be used' to actual output and economic value.
Source: Fast Technology / Iluvatar CoreX media open day (paraphrased and translated from Chinese technology-media reporting; reprinted for informational purposes with attribution).
