According to recent reports, at Iluvatar CoreX's "From Model to World 2026" Media Open Day held on October 28, its next-generation AI chip, the Tiangai 300, was demonstrated live to the media for the first time.
Live demonstrations of the Tiangai 300 were available at the event, accepting comparative tests against the H200 in scenarios involving large language models like GLM-5.3, text-to-image, and text-to-video. The Tiangai 300 outperformed the H200 overall.
In terms of training, during an 8-card full-machine FP8 precision training demonstration based on the Qwen3-32B model, the effective throughput per card of the Tiangai 300 reached 1.33 times that of the H100.
The Tiangai 300 was officially released at the World Artificial Intelligence Conference this July. It adopts the Tianshu Tianshu architecture, integrates next-generation Tensor Cores, and is equipped with 144GB of VRAM. It natively supports three types of computation modes: scalar, vector, and tensor, covering various data precisions such as FP4, FP8, FP16, and BF16, and supports high-speed inter-chip connectivity.
Third-party actual measurements completed by Shanghai INESA this July showed that under typical PD-separated deployment for large models, the overall end-to-end throughput per converted card of the Tiangai 300 reached 1.05 times that of the H200. This marks the first domestic general-purpose GPU whose measured performance reaches the level of the H200.
Further testing by Iluvatar CoreX's Performance Laboratory covers typical business scenarios:
With a 64K long context, the Attention performance of the Tiangai 300 is 1.05 times that of the H200;
For the GLM-5.3 model with a 64K scale, while ensuring the same user experience, the MoE throughput is 1.15 times that of the H200;
In scenarios with long contexts ranging from 8K to 128K, the Agent's first-token latency is reduced by approximately 10% to 30% compared to the H200;
Multimodal generation inference throughput is increased by approximately 2% to 5% compared to the H200;
In online services such as search, advertising, and recommendations, recommendation inference throughput increases by 21%, and lightweight computation latency decreases by 26%.
On the scale-up delivery side, the Tiangai 300 has formed a complete system solution ranging from accelerator cards and 8-card servers to supernodes. The Tiangai 300 IXL144 supernode integrates 144 GPUs within a single standard rack, offering approximately 10.1TB of VRAM. The cost per million effective tokens can be reduced by about 40% compared to mainstream international solutions.