Iluvatar CoreX reports Tiangai 300 inference results against Nvidia H200

The third-party benchmark covered four mainstream models and three application scenarios[3]. In a typical deployment that separates prefill and decode, the Tiangai 300 delivered 1.05 times the H200’s single-card end-to-end throughput[3]. The test also recorded 1.05 times the H200’s attention performance in a 64K long-context scenario[3]. For GLM-5.3 at 64K context under the same user-experience requirements, its MoE throughput was 1.15 times the H200’s[3]. The 1.05-times throughput result is tied to specific test conditions and does not establish performance across all tasks[3].

In 8K-to-128K long-context scenarios, agent first-token latency was about 10% to 30% lower than the H200’s[3]. Multimodal-generation inference throughput increased by about 2% to 5%[3]. Iluvatar also reported that the Tiangai 300 outperformed the H200 on complex text-to-image and text-to-video workloads using major language models, including Zhipu AI’s GLM-5.3[2]. For online search, advertising and recommendation workloads, reported recommendation inference throughput rose 21%, while lightweight-computing latency fell 26%[3]. These figures describe different tests and should not be treated as one broad performance measure[3].

The open day also showcased Iluvatar’s AFD separation solution, IXL144 supernode, computing software stack and Tongyang edge-computing terminal[1]. Iluvatar says AFD completed deployment testing and reduced decode-inference latency by 60%[3]. The IXL144 combines 144 GPUs and about 10.1 terabytes of aggregate memory in a standard rack[1]; Iluvatar estimates its cost per million effective tokens is about 40% lower than international mainstream solutions[3]. The platform supports mixed training and prefill-decode or AFD separation, and can work with other vendors’ accelerators and specialized chips[1].

The H200 is the most advanced Nvidia AI processor approved for export to China under current US trade rules, according to the South China Morning Post[4]. Iluvatar began listing in Hong Kong in January[2]. Analysts said the Tiangai 300 results point to Chinese domestic GPU makers moving from early viability toward enterprise-grade production for large-scale commercial AI deployments[2]. That is an assessment of the broader industry direction, not proof that the reported benchmark results will hold for every customer workload[3][2].