Iluvatar CoreX's Tiangai 300 Benchmarked Against H200 in Real Tests, and On-Device Terminal Tongyang TY1200 Released

“Do not let the rigidity of compute limit the evolution of algorithms. Compute should not become the shackles that constrain algorithms; it should become the solid foundation for incubating new algorithms.” This statement by Shan Tianyi, head of AI and accelerated computing at Iluvatar CoreX, captures this domestic GPU company’s understanding of “high-quality compute” and also serves as a footnote to its challenge to NVIDIA’s H200.

On October 28, the Iluvatar CoreX 2026 Media Open Day was held in Shanghai. The Tiange 300 was opened to the media for a live demonstration for the first time, with side-by-side comparisons against the H200 on tasks such as large language models, text-to-image, and text-to-video, covering scenarios including complex model inference, Token factories, and physical AI. According to actual tests by an authoritative third-party institution, across four mainstream models and three types of scenarios, under a typical PD-disaggregated deployment for large models, the Tiange 300’s normalized single-card overall end-to-end throughput reached 1.05 times that of the H200, making it the first domestic general-purpose GPU to reach the H200 level in actual tests. This figure compares not theoretical peak performance, but the actual output capability of the entire inference chain under specific models and deployment conditions.

The H200 is a representative product of NVIDIA’s Hopper architecture, equipped with 141GB of HBM3e memory and 4.8TB/s of bandwidth. In late 2025, it received approval from the U.S. government to return to the Chinese market, with a 25% fee payable to the U.S. side for each chip exported. As the H200 legally returns to the Chinese market and competes on the same stage with domestic GPUs, this is no longer merely a “behind-closed-doors benchmarking exercise,” but a test against the same yardstick in real business scenarios. The focus of competition for domestic GPUs has accordingly shifted from solving the question of “whether it can be used” to competing on actual output, operating efficiency, and economic value.

Faced with the high wall built by NVIDIA through the CUDA ecosystem and generational iteration, where does a domestic GPU company founded more than a decade ago draw the confidence to directly confront the H200 and even later products?

Turning “benchmarking” into reproducible actual tests

The reason the H200 has become a benchmark is that it is a representative high-performance GPU in the fields of large-model inference and complex computing. For domestic GPUs, reaching this level means crossing a performance threshold of practical significance.

In addition to testing by an authoritative third-party institution, data from the Iluvatar CoreX performance laboratory further covers multiple typical business scenarios: under a 64K long context, the Tiange 300’s Attention performance is 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 is 1.15 times that of the H200; in 8K to 128K long-context scenarios, Agent first-token latency is reduced by about 10%–30% compared with the H200; multimodal generation inference throughput is improved by about 2%–5% compared with the H200; and in online businesses such as search, advertising, and recommendation, recommendation inference throughput is increased by 21% and lightweight computing latency is reduced by 26%.

Behind these metrics is an experience users can perceive: long context determines 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 relates to the wait time after a user sends a request. This means the Tiange 300’s benchmarking is not a “peak contest” in the laboratory, but places performance into real business chains.

The on-site live demonstration moved this benchmarking from data to something perceptible. Shan Tianyi compared the performance of the Tiange 300 and the H200 in text-to-image and text-to-video on site, with the former outperforming the latter in both. In the text-to-image segment, when media at the event specified the generation of an image of “a girl with flaxen hair wearing a white dress walking on ice, cyberpunk style,” the Tiange 300 generated the first image in 4 seconds. In the text-to-video segment, another media outlet asked the products to generate videos of “Doubao wearing knight armor to marry Nailong”; the Tiange 300’s throughput metric was 0.628, higher than that of the H200, and the task completion speed and server-side inference speed were faster. Opening the comparison segment directly to media on-site prompts was itself the best presentation of confidence in the actual test results.

On the training side, the Tiange 300 also provides a reference. In an 8-card full-machine FP8-precision training demonstration based on the Qwen3-32B model, the Tiange 300’s effective throughput per card reached 1.33 times that of the H100. In reinforcement learning, according to conditional engineering estimates, using 1,024 cards to carry out one reinforcement learning run on an existing pretrained base, under the central scenario of 40B activated parameters, a reinforcement learning task of 50 billion generated tokens for a roughly 1T total-parameter MoE model can be completed in about 26 days.

This explains the first source of confidence for the Tiange 300: it does not stand out in just one module, but has entered the range of direct comparison with internationally mainstream high-performance GPUs across multiple types of workloads, including Attention, MoE, Agent latency, multimodality, search/advertising/recommendation, training, and reinforcement learning.

Compute competition becomes a “system ledger”

If 1.05 times is the performance ticket, then the Token factory is the business ledger. As average daily Token calls have grown more than a thousandfold compared with early 2024, the popular form of large-scale compute deployment is shifting from “selling cards” to “selling effective output.” Expenses such as electricity, computer rooms, and fixed operations and maintenance are difficult to compress at will, and unit Token cost depends to a large extent on equipment investment and effective throughput.

Over the past decade, the scale of computing power has grown wildly, but the average utilization rate in inference scenarios is less than 20%, and the average utilization rate in training scenarios is only just over 40% — “theoretical computing power has soared all the way, but when it comes to actual application scenarios, efficiency is greatly reduced.” Shan Tianyi’s analogy was quite blunt: “It’s like driving a car; we face rain, snow, hail, rugged dirt roads, winding mountain roads... It is meaningless to fantasize about running on a flat track. What we need to build is an off-road vehicle that can cut through thorns and move forward bravely even in all kinds of unfavorable scenarios.”

From this, Iluvatar CoreX summarized high-quality computing power with three keywords: high efficiency — the same input produces more effective results; predictability — stable operation over long periods, reproducible performance, and measurable before deployment; sustainability — in the face of model and business iteration, the platform continues to play a role. The confidence of TianGai 300 lies in the fact that it extends single-card performance to system efficiency. Data shows that the step-time fluctuation of TianGai 300 under full load is 0.43%, lower than the 0.76% of the compared Hopper solution. This means it is more stable during long periods of high-load operation, directly safeguarding utilization and availability. For a Token factory, “running fast” is only the first step; “not dropping the ball over long periods” is what determines commercial sustainability.

At the system level, TianGai 300 has formed a complete solution ranging from accelerator cards and 8-card servers to supernodes. The TianGai 300 IXL144 supernode integrates 144 GPUs in a single standard rack and aggregates about 10.1TB of memory. 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 aims to reduce single transmission latency within the domain from 3–5 microseconds to 1 microsecond, and supports unified memory semantic programming and large EP, distributing expert weights across more cores.

Iluvatar CoreX is also trying to open up another possibility with “heterogeneity.” TianGai 300 uses a general-purpose GPU as its base and supports multiple deployment forms such as hybrid training, PD separation, and AFD separation, and can work with accelerator cards and dedicated chips from other vendors. In the AFD separation solution, TianGai 300 handles the full Prefill computation process and Attention in the decoding stage, while partner vendors’ XPUs handle the FFN expert network. Measured Decoding inference latency can be reduced by 60%. It is reported that this is China’s first heterogeneous deployment AFD separation solution based on GPU+XPU.

“The AI computing power system of the future will not necessarily be dominated by a single type of chip.” Shan Tianyi said, “The key is how to organize them efficiently so that the entire system can deliver greater value.” This statement points to the second source of confidence for TianGai 300 in challenging H200: it does not use one chip to handle all computation, but uses a general architecture and heterogeneous collaboration to organize customers’ existing computing power assets into a more efficient system.

Turning expert experience into infrastructure

Once hardware performance reaches a usable range, the real threshold is software and ecosystem, which means whether developers can truly use it well — and this is precisely Nvidia’s deepest moat. New models keep emerging, and deployment, adaptation, and performance tuning often require experienced engineers to invest a great deal of time. If the software stack is complex, hardware performance is difficult to translate into business value.

Iluvatar CoreX’s answer is to bring AI into the development of the software stack itself. Shi Jiasheng, Vice President and head of software development, for the first time demonstrated the layout of Iluvatar CoreX’s software stack fully embracing Agentic AI, unfolding around three levels: “easy to use, easy to optimize, easy to develop,” combining the software stack, professional tools, and engineering experience with Agent technology, so that frontline developers can directly call tools, reuse experience, and carry out optimization around scenario goals and service constraints.

In terms of “easy to use,” in measured model PD-separated deployment, the overall time for environment preparation and service deployment was shortened from 1 hour to 10 minutes, and manual operations were reduced from dozens of steps to 1 step. In terms of “easy to optimize,” according to a measured tuning task for a large model, the work cycle was shortened from 7 days to within 2 hours; under the same hardware, model, and precision requirements, the running performance of related tasks nearly doubled compared with before optimization. In terms of “easy to develop,” as the first domestic GPGPU vendor to deeply integrate FlyDSL, Iluvatar CoreX uses Agent and FlyDSL collaboration to let AI participate in code generation, compilation, testing, and performance optimization; in a measured adaptation and tuning task for a mainstream model, work that previously took 10 person-days can now be completed by 1 person within hours, and compilation speed has increased by more than 90%.

“Agentic AI allows one engineer’s experience to continuously serve thousands of developers.” Shi Jiasheng said. For domestic GPUs, this is not only a matter of tool efficiency, but also an ecosystem issue. Precipitating the methods accumulated by GPGPU engineers in model adaptation and performance tuning into callable tools and skills, and turning vendors’ scarcest asset — expert experience — into infrastructure that customers can reuse, is a further answer than “compatibility” to the existing inertia of the CUDA ecosystem: rather than forcing developers to migrate reluctantly, it is better to make migration and tuning themselves cheap enough.

Only when the thresholds for deployment, tuning, and Kernel development are lowered can hardware capabilities enter more models, more businesses, and more developers’ hands more quickly. This is the third source of confidence for TianGai 300 in challenging H200 — using the software stack to productize “expert experience” and turn point performance into replicable productivity.

“Folding” cloud capabilities into a tiny space

Beyond TianGai 300, Iluvatar CoreX is also filling in another piece of the cloud-edge-device puzzle. Its Tongyang edge-side compute terminal TY1200, rated at 300 TOPS dense compute, surpassed NVIDIA's currently very strong edge product Thor T5000 in multiple sets of actual measurements.

The Tongyang TY1200 is rated at 300 TOPS INT8 dense compute, with a storage configuration of 32GB system memory plus 32GB HBM2e video memory; its comparison target, NVIDIA Thor-T5000, has 2P FP4 sparse compute and 128GB shared memory. Materials show that in actual tests of LLM models such as Qwen3.8-27B and Qwen3.6-35B-A3B, the TY1200's performance reached 1.2 times that of Thor T5000; in Cosmos-R2-2B and Cosmos-R2-8B tests, the TY1200's measured performance reached 158% and 180% of Thor T5000, respectively.

The parameter bases are not the same, but at the event site, Song Yuanying, marketing head of Iluvatar CoreX's Tongyang products, likewise conducted actual measurement comparisons from real tasks. When the media asked the Qwen3.8-27B model “who would win if a lion and a tiger fought one-on-one,” the TY1200 had lower first-token latency, higher throughput, and shorter output time. In a text-to-video task, the media asked the two products to each generate a video of “cats fighting and breaking a vase,” and the TY1200 generated a 5-second video within 6 minutes and 10 seconds. With smaller memory specifications and a dense-compute basis, achieving faster responses in actual measured tasks is precisely why Iluvatar CoreX repeatedly emphasizes “not being defined by rated compute.”

“For a cloud GPU vendor to do edge-side products is not about making chips smaller, but about pushing capabilities downward,” Song Yuanying said. Iluvatar CoreX brings the model compression, quantization, distillation, and software-hardware co-design capabilities accumulated in large-scale cloud-side training and inference, together with its general-purpose GPU architecture, compiler, operator libraries, and performance tools, to the edge, so that models are “born in the cloud and grow at the edge.”

Behind this is the realistic logic of compute tiering. “The closer compute is to data, the faster value is realized,” Song Yuanying emphasized in the speech, and the upper limit of edge-side competitiveness is determined by cloud-side full-stack AI capabilities. High-confidentiality scenarios such as contract review and investment BP analysis require that “data does not leave the premises,” physical AI and embodied intelligence require millisecond-level response, simple tasks do not all need to go to the cloud, and complex tasks should not all be pushed to the edge either. The deployment of the TY1200 in scenarios such as energy-storage battery cell research, emergency visual recognition, cultural tourism guidance, and companion robots for medical visits shows that the edge side is not a “shrunk-down version of the cloud,” but a new site for high-quality compute.

A telling detail at the event was that when replicating the MicroDuck robotic duck scenario demonstrated by NVIDIA at GTC Berlin, Iluvatar CoreX did not, like NVIDIA, send the training task to a cloud supercomputer; instead, it completed all physical simulation, policy training, and motion verification on the offline, network-disconnected TY1200, ultimately enabling a 3D-printed, hand-assembled “handmade version” robotic duck to stand up successfully. Song Yuanying said this was the first time GPU-accelerated physical simulation based on MuJoCo-Warp had been achieved on a domestic edge-side product.

Conclusion: Crossing the line is a new starting point

From the cloud-side TianGai 300 and the Iluvatar CoreX computing software stack to the edge-side Tongyang compute terminal, Iluvatar CoreX is building an enhanced loop of cloud-edge-device collaboration. This loop allows high-performance computing to go beyond data centers and enter the physical world and industry sites. For Iluvatar CoreX, the confidence to challenge H200 comes not only from a single chip, but also from the capability transfer implied by “however deep the cloud is, the edge can go that far.”

This confidence is essentially a set of system capabilities: organizing chip performance, software efficiency, heterogeneous organization, and scenario deployment into effective output that can be delivered sustainably. The actual test questions randomly given by media at the event are the most intuitive footnote to this set of system capabilities. What TianGai 300 crossed is not just a performance threshold, but a key milestone in domestic general-purpose GPUs moving toward competition in high-value applications.

Crossing the line is even more a new starting point. It is believed that domestic GPUs, including Iluvatar CoreX, will move from “whether it can be used” to “whether it is good to use and whether it can be used at scale,” using customers' real workloads as the touchstone, continuously verifying and realizing value in sustained delivery and long-term application, opening up greater market space, and writing a new chapter in the high-quality development of domestic compute.