Iluvatar CoreX Tianyao 300 live demonstration benchmarks against H200, and previews 600T on-device product

On-site random questions were posed, such as "a tiger kicks the moon away," and both a domestic GPU and NVIDIA's flagship high-end inference card H200 were asked to answer. The domestic GPU took 95.6 seconds, while the NVIDIA H200 took 112 seconds. This GPU that beat the H200 is called Tiangai 300. At a recent "From Models to the World" media open day hosted by Tianshu Zhixin, one of the "Four Little Dragons of Shanghai GPUs," the Tiangai 300 was opened to the media for a live demonstration for the first time. In July of this year, third-party testing completed by an authoritative third-party organization showed that under typical PD-disaggregated deployment, the Tiangai 300's converted single-card overall end-to-end throughput reached 1.05 times that of NVIDIA's flagship high-end inference card H200. The AFD disaggregation solution based on the Tiangai 300, the IXL144 supernode solution, as well as the Tianshu intelligent computing software stack and the Tongyang edge-side computing terminal, appeared on the same stage, and a next-generation edge-side product with a nominal computing power of 600T was previewed. Tianshu Zhixin summarizes this path as producing intelligence in the cloud, accelerating adaptation and optimization through software, and enabling intelligence to enter more on-site scenarios at the edge, forming a reinforcing loop that allows high-quality computing power to grow continuously. Benchmarking against H200 in actual tests: the performance threshold for moving toward complex AI tasks. In July of this year, an authoritative third-party organization conducted tests on four mainstream models and three types of scenarios. Under the typical PD-disaggregated deployment of large models, the Tiangai 300's converted single-card overall end-to-end throughput reached 1.05 times that of the H200. What is compared here is not the theoretical peak, but the actual output capability of the entire inference chain under specific models and deployment conditions. Actual tests by the Tianshu Zhixin Performance Lab further covered typical business scenarios: under a 64K long context, the Tiangai 300's Attention performance was 1.05 times that of the H200; under the GLM-5.3 model at the 64K scale, while ensuring the same user experience, MoE throughput was 1.15 times that of the H200; in 8K to 128K long-context scenarios, Agent first-token latency was reduced by about 10%-30% compared with the H200; multimodal generation inference throughput increased by about 2%-5% compared with the H200; and in online businesses such as search, advertising, and recommendation, recommendation inference throughput increased by 21% and lightweight computing latency decreased by 26%. The event site opened a live demonstration of the Tiangai 300 and accepted comparative tests with the H200 in scenarios such as the GLM5.3 large language model, text-to-image, and text-to-video. These metrics each correspond to an experience that users can perceive. Long-context processing affects whether long-document analysis and complex Agents can run efficiently; MoE throughput affects the system's ability to handle requests under the same service requirements; and first-token latency directly affects the waiting time after a user sends a request. This benchmarking shows that the Tiangai 300 has entered the range of direct comparison with internationally mainstream high-level GPUs across multiple key workloads. In terms of training, in an 8-card full-machine FP8-precision training demonstration based on the Qwen3-32B model, the Tiangai 300's effective throughput per card reached 1.33 times that of the H100. At present, reinforcement learning has become a core post-training route publicly adopted by major models at home and abroad, including OpenAI, Google, Meta, xAI, as well as DeepSeek, Qwen, Minimax, and Kimi. The Tiangai 300 is equally capable of reinforcement learning training. According to conditional engineering estimates, using 1,024 cards to carry out one round of reinforcement learning on an existing pretrained base, under the central scenario of 40B activated parameters, a reinforcement learning task of 50 billion generated tokens for a MoE model with about 1T total parameters can be completed in about 26 days. As average daily Token calls have grown more than a thousandfold compared with early 2024, the Token factory is becoming a popular form for the large-scale implementation of computing power, and is also a concentrated testing ground for the three elements of high-quality computing power. Various performance indicators show that the Tiangai 300 highly meets the practical requirements of Token factories for computing power. The first is high efficiency. Expenditures such as electricity, computer rooms, and fixed operations and maintenance are difficult to compress at will, and the cost per Token depends to a large extent on equipment investment and effective throughput. Under the aforementioned third-party test conditions, the Tiangai 300's converted single-card end-to-end throughput reached 1.05 times that of the H200, providing a performance foundation for producing more effective Tokens per unit time. The second is predictability. A Token factory cannot only run fast in short-term tests; it must also remain stable during long periods of full-load operation. The Tiangai 300's step-time fluctuation under full-load operation was 0.43%, lower than the 0.76% of the compared Hopper solution. It does not falter under long periods of high load, directly safeguarding utilization and availability. Finally, sustainability. The Tiangai 300's general architecture natively supports CUDA semantics, mainstream frameworks have all been adapted, and it provides different deployment forms from accelerator cards and servers to supernodes, oriented toward a variety of models and business needs. The Tiangai 300 also provides an open, composable heterogeneous computing platform solution. With a general GPU as the base, it supports multiple deployment forms such as hybrid training, PD disaggregation, and AFD disaggregation, and can collaborate with accelerator cards and dedicated chips from other vendors. Taking AFD disaggregation as an example, the Tiangai 300 handles the full Prefill computation process and Attention in the decoding stage, while partner vendors' XPUs handle the FFN expert network. The two types of chips collaborate heterogeneously under unified scheduling, and actual tests show that Decoding inference latency can be reduced by 60%. This solution has been actually run through and is China's first heterogeneous deployment AFD disaggregation solution based on GPU+XPU. From "whether it can be used" to competing on actual output, operating efficiency, and economic value. "The AI computing systems of the future may not be dominated by a single type of chip."Shan Tianyi, head of AI and accelerated computing at Iluvatar CoreX, said, "The key is how to organize them efficiently so that the entire system can deliver greater value. We hope that through open heterogeneous computing capabilities, customers can fully leverage their existing computing assets and build computing systems suited to their own business in a more flexible way." The value of high-quality computing power thus lies not merely in test scores, but in the ability to be continuously committed to actual production. Iluvatar CoreX further extends general-purpose computing into frontier fields such as scientific intelligent computing and physical AI, accelerating scientific discovery and industrial applications. The Tiange 300 fully supports models such as Boltz-2, OpenFold3, and Protenix as well as frameworks such as Newton, Warp, and ALCHEMI, covering nearly a hundred typical scenarios across four major fields—life sciences, physical AI, chemistry, and materials—including key tasks such as protein structure prediction, drug screening, robot simulation, molecular dynamics, and material structure optimization. Physical AI in particular demonstrates the new value of general-purpose computing. For robots to enter the real world, they often need to repeatedly trial and error in simulation environments that follow physical laws, learning actions and validating strategies. The cloud-based simulation and computing capabilities of the Tiange 300 make such learning and validation more thorough. On the large-scale delivery side, the Tiange 300 has formed a complete system solution ranging from accelerator cards and 8-card servers to super nodes. The Tiange 300 IXL144 super node integrates 144 GPUs within a single standard rack and aggregates approximately 10.1TB of memory, making it the industry's highest-density single-rack GPU super node at present. According to estimates, its cost per million effective tokens can be reduced by about 40% compared with mainstream international solutions. Its new-generation P2P interconnect engine is designed for both Scale-up and Scale-out, with the goal of reducing single intra-domain transmission latency from 3–5 microseconds to 1 microsecond; it supports unified memory semantic programming; and it also supports large EP, distributing expert weights across more cores. This solution has already been publicly unveiled at the Industrial Fair, and other super node solutions for different computing scales and data center conditions will also be launched successively. Tongyang TY1200 Outperforms Thor T5000 in Actual Tests: Cloud Capabilities Pushed Down to the Device Side On the day of the open day, when the on-site software stack demonstrated a new path of "using AI to develop AI," the Tongyang TY1200 surpassed Intel Thor T5000 in actual tests—a highly capable high-performance edge computing platform from Nvidia. In a demonstration where people on site were randomly selected to ask questions, the Tongyang TY1200's response speed clearly beat the Thor T5000. The Tongyang TY1200 edge computing terminal is rated at 300 TOPS dense computing power, rather than the industry's customary sparse peak; in actual tests with LLM models such as Qwen3.8-27B and Qwen3.6-35B-A3B, the TY1200's performance reached 1.2 times that of the Thor T5000. With measured dense computing power surpassing chips with higher rated computing power, the effective output at the edge is likewise not defined by rated computing power. "For a cloud GPU vendor to do edge devices is not about making chips smaller, but about pushing capabilities down," said Song Yuanying, head of product marketing for Tongyang at Iluvatar CoreX. The model compression, quantization, distillation, and software-hardware co-design capabilities that Iluvatar CoreX has accumulated in large-scale training and inference on the cloud side, together with its general-purpose GPU architecture, compiler, operator libraries, and performance tools, have been brought to the edge. Models can therefore "be born in the cloud and grow at the edge": trained in the cloud, deployed at the edge, and re-judged by the cloud when needed, so that model capabilities and accuracy remain connected. The high-load delivery experience accumulated at customer sites also provides a capability foundation for delivery in scenarios with high stability requirements, such as industrial applications and embodied intelligence. Tongyang covers needs ranging from customer self-integration to ready-to-use deployment with one ecosystem, two forms, and five products; the device-cloud continuity technical approach, "folding cloud capabilities into a tiny space," also gives customers' edge investments long-term continuity. The TY1200 previously won the 2026 Industrial Fair CIIF Guoxin Exhibition Innovation Achievement Award. At the event, Iluvatar CoreX also previewed its next-generation edge product, the Tongyang TY2000, which will have a rated computing power of 600T, pushing the boundary of pushing cloud capabilities down to the edge one step further.