Iluvatar CoreX First Live Demonstration of TianGai 300 and On-Device Terminal TY1200

ChipInside reported on October 28 that today, Iluvatar CoreX demonstrated for the first time on real hardware its general-purpose GPU Tiangai 300 and the edge computing terminal TY1200. It also previewed its next-generation edge product TY2000, with a nominal computing power of 600T, supporting multi-machine interconnection, elastic expansion, and offline deployment.

Iluvatar CoreX disclosed that third-party testing completed by Shanghai Yidian in July this year showed that under a typical PD-disaggregated deployment, the Tiangai 300's converted single-card overall end-to-end throughput reached 1.05 times that of the H200, making it the first domestic general-purpose GPU whose measured performance reaches the level of NVIDIA's H200. In addition, its edge computing terminal Tongyang TY1200 also surpassed NVIDIA's Thor T5000 in actual testing.

Shan Tianyi, head of AI and accelerated computing at Iluvatar CoreX, said that AI is moving from information intelligence to action intelligence, and then to exploration intelligence, forming a flywheel of intelligence. Information intelligence is the ability to understand, generate, and transform information; action intelligence gives AI the ability to act, to plan, execute, and provide feedback on its own; exploration intelligence is AI independently designing, validating, and iterating, such as in fields like AI for Science, physical AI, and embodied intelligence.

In January this year, the company announced its four-generation chip architecture roadmap: the Tianshu architecture was released in 2025, the Tianxuan architecture in 2026, the Tianji architecture in 2026, and the Tianquan architecture is planned for release in 2027.

What Iluvatar CoreX does is provide a continuous supply of computing power for each cycle of the intelligence flywheel mentioned above.

1. TY1200's measured performance exceeds Thor T5000; edge computing power trains a bipedal duck robot

On this basis, Song Yuanying, head of marketing for Iluvatar CoreX's Tongyang products, said that the cost-effectiveness of computing power is currently a key focus of the industry, and whoever can offer the best cost-effectiveness can seize the initiative.

An obvious trend behind this is that computing power needs to be tiered, which is the core reason for Iluvatar CoreX's layout of edge-side products.

In January this year, Iluvatar CoreX officially released its edge series Tongyang and formed a 1+2+5 matrix.

1 refers to an ecosystem composed of a mature general-purpose GPU architecture and a complete software stack; 2 refers to two product forms, computing modules and computing terminals, provided to two types of customers; and 5 refers to five products, among which the computing module products include TY1000 and TY1100, and the computing terminal products include TY1100-NX, TY1100-NX-PRO, and TY1200.

Song Yuanying conducted an on-site measured comparison of the effects of TY1200 and NVIDIA Thor-T5000. Regarding the parameter comparison of the two devices, Thor-T5000 has 2P FP4 sparse computing power and 128GB of shared memory; TY1200 has 300 TOPS INT8 dense computing power, with storage consisting of 32GB system memory + 32GB HBM2e video memory.

They chose the Qwen3.8-27B model and first asked the model, "Who would win in a one-on-one fight between a lion and a tiger?" The results showed that TY1200 had lower first-token latency, higher throughput, and shorter output time.

According to data publicly released by Iluvatar CoreX, in actual tests of large language models, multimodal models, world models, and others, TY1200's TPS (tokens per second) all surpassed that of T5000. Song Yuanying said the reason for TY1200's better performance is higher video memory bandwidth and a greater number of compute cores.

The next task tested on-site by the demonstrator was text-to-video. She had the two products generate a video of cats fighting and breaking a vase. TY1200 generated a 5-second video within 6 minutes and 10 seconds.

Next was a legal compliance scenario with higher confidentiality requirements, where the demonstrator on-site demonstrated legal contract review based on internal corporate rules.

Finally was the physical AI scenario. They trained the bipedal duck robot MicroDuck on-site using edge computing power, enabling MicroDuck to complete various activities in a simulated environment, including moving forward, backward, turning, and standing up after being pushed down.

The demonstrator also had the physical MicroDuck walk on its own on-site and stand up after falling.

In terms of specific applications, they built a vertical AI agent for energy storage battery cell research based on TY1200; in addition, in the field of visual recognition, the recognition accuracy of traditional solutions is 70%. Iluvatar CoreX used CV + multimodal + edge computing power, combined with the cloud for emergency command management, to raise visual recognition accuracy to 95%.

In addition, a robot dog based on the TY1000 computing module can automatically recognize and follow its owner, perceive, decide, and act, with the entire process completed locally; a patient accompaniment robot built on the TY1200 computing terminal can guide patients and understand hospital visit procedures.

2. The Tiangai 300 born for next-generation AI; on-site testing had "Doubao marry Nailong"

Shan Tianyi highlighted the new computing platform Tiangai 300 built for next-generation AI. The Tiangai 300 has 144GB of video memory, supports high-speed inter-chip interconnection, is equipped with a new-generation Tensor Core, and supports data precisions such as FP4, FP8, FP16, and BF16.

Actual testing by the Iluvatar CoreX Performance Lab shows that under a 64K long context, the Tiangai 300's Attention performance is 1.05 times that of the H200.

With the GLM-5.3 model at the 64K scale, while ensuring the same user experience, the Tiangai 300's MoE throughput is 1.15 times that of the H200.

Under 8K to 128K long-context scenarios, the Tiangai 300's Agent first-token latency is about 10%–30% lower than the H200's; in multimodal generation inference throughput, the Tiangai 300 improves on the H200 by about 2%–5%; in online businesses such as search, advertising, and recommendation, recommendation inference throughput increases by 21% and lightweight computing latency decreases by 26%.

On the training side, 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.

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. According to conditional engineering estimates, when the Tiangai 300 uses 1,024 cards to carry out one round of reinforcement learning on an existing pretrained base, under the central scenario of 40B activated parameters, it can complete a 50-billion-generated-token reinforcement learning task for a roughly 1T-total-parameter MoE model in about 26 days.

Shan Tianyi conducted on-site actual testing of the Tiangai 300 and H200 in text-to-image and text-to-video performance.

It generated an image of "a girl with flaxen hair wearing a white dress walking on ice, cyberpunk style," and the Tiangai 300 generated the first image in 4 seconds.

In addition, it generated on site a video of "Doubao wearing knight's armor marrying Nailong." The Tiangai 300's throughput was 0.628, higher than the H200's, and the task completion speed and server-side inference speed were faster.

At the level of large-scale delivery, Shan Tianyi revealed that the Tiangai 300 has formed a complete system solution from accelerator cards and 8-card-per-machine servers to super nodes. The Tiangai 300 IXL144 super node integrates 144 GPUs in a single standard rack and aggregates about 10.1TB of video 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 aimed at Scale-up and Scale-out, with the goal of reducing a 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.

3. The first domestic GPGPU vendor to deeply integrate FlyDSL, with compilation speed increased by over 90%

Shi Jiasheng, Vice President of Iluvatar CoreX and head of software development, showcased the systematic layout of the Iluvatar CoreX software stack for Agentic AI around three aspects: ease of use, ease of optimization, and ease of development.

First is ease of use. The Iluvatar CoreX software stack organizes environment installation, configuration checks, and model deployment methods into tools and skills that agents can call, with AI assisting execution and checking results, reducing the burden on developers of repeatedly consulting documentation, manually configuring, and troubleshooting problems.

In an actual test of model PD-disaggregated 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.

Second is ease of optimization. The software stack combines self-developed performance analysis tools and tuning experience to let AI locate bottlenecks based on real runtime data, propose optimization plans, and verify the effects through testing. According to an actual test of a certain large-model tuning task, the work cycle was shortened from 7 days to within 2 hours; under the same hardware, model, and precision requirements, the relevant task's running performance nearly doubled compared with before optimization.

Finally, ease of development. He mentioned that Iluvatar CoreX is the first domestic GPGPU vendor to deeply integrate FlyDSL, and through the collaboration of agents and FlyDSL, it lets AI participate in code generation, compilation, testing, and performance optimization.

Shi Jiasheng revealed that in an actual test of adaptation and tuning for a certain mainstream model, work that previously took 10 person-days can now be completed by 1 person within hours, with compilation speed increased by more than 90%; in an ideal state, 1% of compilation time can be used to achieve 99% of runtime performance.

Conclusion: Cloud plus software collaboration; domestic computing power is moving toward an autonomous evolution stage

The Tiangai 300 reached the level of comparable mainstream international products in real business tests, marking that domestic general-purpose GPUs are beginning to have the strength to develop rapidly in complex AI tasks and the high-computing-power market.

From its layout perspective, the cloud efficiently produces intelligence, software accelerates evolution, and the device side broadens real-world applications and connects to the physical world. The collaboration of the three ends allows efficient, predictable, and sustainable computing power to accumulate in a cycle. Models continue to become stronger in the cloud and sink down to the device side for implementation; real scenarios on the device side in turn feed back into cloud optimization of operators, quantization strategies, and model structures, continuously lowering implementation costs and improving business stability, ultimately forming a positive flywheel of autonomous evolution for domestic computing power.