Iluvatar CoreX debuts Tiangai 300 GPU with H200-level claims
Iluvatar CoreX publicly demonstrated the Tiangai 300 general-purpose GPU for the first time at its Shanghai media open day, alongside the IXL144 supernode, TianSuan compute software stack and T-Centric edge products. The company and media reports cite third-party and internal tests showing 1.05 times NVIDIA H200 per-card throughput or Attention performance under specified conditions, while also claiming a 60% latency reduction for a GPU+XPU AFD deployment. The company also previewed the T-Centric TY2000, saying it would offer 600T of compute, without disclosing detailed specifications, power consumption, pricing or availability.
The first live demonstration provides a new deployment reference for a domestic AI GPU, but the headline performance and cost comparisons remain vendor-reported and depend on workloads and configurations that have not been fully disclosed.
10月28日，天数智芯首次开放天垓300通用GPU实机演示，展示基于天垓300的AFD分离方案、IXL144超节点、天数智算软件栈及彤央端侧算力终端，并预告下一代端侧产品TY2000，标称算力达600T。
据天数智芯披露的第三方测试结果，在典型PD分离部署下，天垓300折算单卡端到端吞吐达到NVIDIA H200的1.05倍。公司内部测试显示，该产品在部分长上下文、混合专家模型（MoE）和在线推荐场景中也取得了相应的性能表现。端侧方面，彤央TY1200在指定大模型测试中的性能达到NVIDIA Thor T5000的1.2倍。
天垓300：部分推理场景达到H200水平
据天数智芯介绍，第三方机构针对四种主流模型、三类应用场景进行了测试。在典型PD分离部署下，天垓300折算单卡端到端吞吐达到H200的1.05倍。PD分离将大模型推理中的预填充（Prefill）和解码（Decode）阶段分开部署，实际性能取决于模型、并发规模、计算与通信开销等因素，因此这一结果对应特定测试条件，不代表所有任务中的性能表现。
#Iluvatar CoreX#Tiangai 300#IXL144#TianSuan compute software stack#T-Centric TY1200#T-Centric TY2000#NVIDIA H200#NVIDIA Thor T5000#Qwen3-32B#Qwen3.8-27B#GLM-5.3#MuJoCo-Warp