AnswersPublished15 sources
Iluvatar CoreX’s Tiangai 300 Demo: A 1.05× H200 Claim, With Key Details Still Undisclosed
Iluvatar CoreX showed the Tiangai 300 running on hardware for the first time and cited third party testing that put its single card throughput at 1.05× NVIDIA H200 in a specific deployment. The event also featured an IXL144 rack, GPU–XPU inference disaggregation and the TY1200 edge system, with performance and cost...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Iluvatar CoreX showed the Tiangai 300 running on hardware for the first time and cited third party testing that put its single card throughput at 1.05× NVIDIA H200 in a specific deployment.
- The event also featured an IXL144 rack, GPU–XPU inference disaggregation and the TY1200 edge system, with performance and cost figures presented as company claims.[3][4]
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Iluvatar CoreX demonstrate and claim at its October 28, 2026 media open day in Shanghai about the Tiangai 300 GPU’s first live hard. Article summary: At its October 28, 2026 Shanghai media open day, Iluvatar CoreX publicly demonstrated the Tiangai 300 running on real hardware for the first time, alongside heterogeneous inference and edge-AI systems—a meaningful move b. Topic tags: general, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake numbers, clic
Iluvatar CoreX’s October 28, 2026 media open day in Shanghai moved the Tiangai 300 from architecture and roadmap presentations to a public live-hardware demonstration. The company also presented a supernode, a heterogeneous inference setup and an edge-computing system. Its headline claim—that third-party testing put Tiangai 300 throughput at 1.05 times NVIDIA’s H200—needs context: the reported result applies to a particular deployment, and key test details were not disclosed.2
3
What Iluvatar CoreX demonstrated
The event put the Tiangai 300 general-purpose GPU on working hardware in front of media for the first time, alongside the company’s software stack, IXL144 supernode and Tongyang TY1200 edge-computing product.3
6 That is a practical step beyond showing an architecture or product roadmap, though a live demonstration alone does not verify comparative performance claims.
What the 1.05× H200 figure means
Iluvatar said third-party testing found that Tiangai 300’s normalized, single-card overall end-to-end throughput reached 1.05× NVIDIA H200’s in a typical prefill/decode (PD) disaggregated deployment. The company described it as the first domestic general-purpose GPU to reach H200-level measured performance.2
3
That figure describes a reported result for a particular setup—not a finding that the Tiangai 300 is 5% faster on every model, task or configuration. A report covering the announcement noted that test methods, model loads and workload conditions had not been released alongside the claim.3 The reported third-party test is relevant evidence, but the missing detail makes it difficult for others to reproduce or assess the comparison.
IXL144: a rack-scale deployment claim
Iluvatar also discussed the IXL144 supernode, which had been shown at an earlier industrial exhibition. Its stated single-rack configuration is 144 GPUs, 18 CPUs and about 10.1 TB of GPU memory.3
The company described the system as having industry-leading GPU density in a single rack and estimated that it could reduce cost per million effective tokens by 40% compared with mainstream international solutions.3
9 Those are company claims, and the cost figure is an estimate—not a published, independently verified operating-cost comparison.
AFD disaggregation: splitting work between accelerators
The company said it had successfully run a domestic heterogeneous GPU–XPU AFD deployment and reported a 60% reduction in inference latency.3 In this approach, parts of a model’s work—such as attention and feed-forward network computations—are assigned to different accelerator types. Reporting on the setup said the Tiangai 300 handled the prefill process and decode-stage attention, while a partner XPU handled the feed-forward network.
4
The result is tied to that reported deployment. Without disclosed test conditions and a precise account of how latency was measured, it should not be treated as a universal improvement for other models or systems.3
4
TY1200 brings the claims to edge AI and robotics
The Tongyang TY1200 is specified at 300 TOPS of dense compute and was presented for edge-AI and embodied-intelligence applications.2
3 Iluvatar claimed performance above NVIDIA’s Thor T5000; later reporting described results at about 1.2× T5000 on selected Qwen models. That is a model-specific comparison, not a general peak-performance ranking.
2
The company also said the TY1200 was the first domestic edge product to run GPU-accelerated physics simulation using MuJoCo-Warp. The reported demonstration involved local simulation for robot learning, linking simulation, training and validation on the edge system.2
3 The “first” designation and the wider performance claims remain company-reported assertions in the sources available here.
What the event does—and doesn’t—establish
The event showed more than product plans: the Tiangai 300 ran on hardware, and Iluvatar presented it alongside rack-scale and heterogeneous-computing systems and an edge-AI demonstration.3
6 That gives readers a clearer view of the company’s systems ambitions than an architecture slide alone.
It does not, by itself, establish broad performance superiority over NVIDIA products. The H200 figure is a reported third-party result, but the cited coverage says the test methods, models and workload conditions were not disclosed.3 A stronger comparison would publish the model and version, precision, batch and input/output settings, hardware configuration, software stack, latency definition and, for cost claims, the assumptions behind the calculation. Until then, the figures are best read as specific reported results and company estimates—not independently reproducible, general-purpose rankings.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Iluvatar CoreX’s Tiangai 300 Demo: A 1.05× H200 Claim, With Key Details Still Undisclosed"?
Iluvatar CoreX showed the Tiangai 300 running on hardware for the first time and cited third party testing that put its single card throughput at 1.05× NVIDIA H200 in a specific deployment.
What are the key points to validate first?
Iluvatar CoreX showed the Tiangai 300 running on hardware for the first time and cited third party testing that put its single card throughput at 1.05× NVIDIA H200 in a specific deployment. The event also featured an IXL144 rack, GPU–XPU inference disaggregation and the TY1200 edge system, with performance and cost figures presented as company claims.[3][4]