TL;DR
Play games on Amazon Luna for your next adventure, included with Prime
- A rotating selection of games, no download needed
- Play on TV, laptop or phone
- Fast, free delivery for your gear, too
Chinese chipmaker Iluvatar CoreX says its Tiangai 300 GPU matched or exceeded NVIDIA’s H200 in selected AI workloads shown in a company-run demonstration. A separate assessment by state-backed Shanghai INESA reported 1.05 times the H200’s throughput under typical deployment conditions, but the testing basis and broader performance remain unclear.
Chinese chipmaker Iluvatar CoreX says its Tiangai 300 GPU can match or outperform NVIDIA’s H200 in selected AI workloads, after showing benchmark results to mainland Chinese media. The figures include a company-run text-to-image test and a separate assessment reporting a 5% throughput edge over the H200, but they do not establish that Tiangai 300 is broadly faster across workloads.
During the live demonstration, Iluvatar ran text-to-image and text-to-video generation using frontier models, including Z.ai’s GLM-5.3. In one image-generation test, the company reported that Tiangai 300 completed all tasks at 0.101 images per second. The H200 completed 87.5% of tasks and produced 0.086 images per second in that comparison. Iluvatar also reported that Tiangai 300 took 0.64 seconds longer to finish the task, a caveat that complicates any simple claim of a clear performance win.
A separate test by Shanghai INESA, a state-backed organization, reportedly measured Tiangai 300 at 1.05 times the H200’s throughput under typical deployment conditions. Wallstreetcn separately reported a 15% Tiangai 300 advantage in mixture-of-experts throughput and up to 30% lower first-token latency in long-text processing. These are distinct reported measurements, not a single standardized result, and the available details do not provide enough information to compare their test conditions directly.
For model training, Tiangai 300 reportedly delivered 1.33 times the effective throughput of NVIDIA’s H100 while fine-tuning Alibaba’s Qwen3-32B. That result is a comparison with the older H100, not the H200. Iluvatar’s demonstration also covered text-to-video generation, but no specific result for that workload was included in the reported figures.
AI ACCELERATORS · CLAIMS UNDER REVIEW
Tiangai 300’s H200 Challenge Comes With a Caveat
Iluvatar CoreX reported wins in selected AI tests. Shanghai INESA separately measured a 5% throughput lead over NVIDIA’s H200 under typical deployment conditions. The results are promising, but they do not establish broad superiority.
IMAGE RATE · TIANGAI
0.101
images per second
IMAGE RATE · H200
0.086
images per second
H200 TASKS
87.5%
completion in company demo
IMAGE TEST CAVEAT
+0.64s
Tiangai took longer to finish
01 / WHAT WAS REPORTED
Different tests, different comparisons
The headline numbers come from separate measurements. Each speaks to a particular workload or setup; together, they do not make one standardized head-to-head result.
COMPANY DEMONSTRATION
0.101 vs 0.086
Text-to-image generation
Iluvatar reported a higher image rate and full task completion for Tiangai 300. H200 completed 87.5% of tasks. Tiangai 300 also took 0.64 seconds longer to finish the task.
SEPARATE ASSESSMENT
1.05×
Typical deployment throughput
Shanghai INESA reportedly measured a 5% Tiangai 300 lead over H200. The available account does not detail enough test conditions to judge how broadly it applies.
OTHER REPORTED FIGURES
15% / 30%
MoE throughput and latency
Wallstreetcn reported a 15% advantage in mixture-of-experts throughput and up to 30% lower first-token latency for long-text processing. These are distinct results.
02 / READING THE IMAGE TEST
A higher rate, with a timing wrinkle
In Iluvatar’s text-to-image demonstration, the company reported a higher images-per-second figure for Tiangai 300. The longer completion time is another measure whose relationship to the rate is not explained in the available report.
Reported image-generation rate, in images per second. Company-run demonstration; test settings and independent repeatability data were not fully disclosed.
03 / A DIFFERENT BENCHMARK
Training result compares against H100
FINE-TUNING · QWEN3-32B
1.33×
Effective throughput versus NVIDIA H100
Tiangai 300 reportedly delivered 1.33 times the effective throughput while fine-tuning Alibaba’s Qwen3-32B. This comparison is with the older H100, not the H200.
“The company says Tiangai 300 can match and, in some cases, beat NVIDIA’s H200 in real-world AI workloads.”
ILUVATAR COREX · COMPANY CLAIM
The demonstration also included text-to-video generation, but the reported account provided no specific performance result for that workload.
04 / MARKET CONTEXT
A small lead in a strategic market
Export limits on advanced accelerators have increased interest in domestic Chinese alternatives. A competitive local GPU could give buyers another option, but performance figures alone do not determine adoption.
CURRENT POSITION
0.3%
Reported domestic GPU share
Iluvatar held 0.3% of China’s domestic general-purpose GPU market in 2024, underscoring the distance between its current position and its ambitions.
COMPETITIVE FIELD
Several local GPU developers
Other named Chinese competitors include Biren, MetaX, Enflame, and Moore Threads. The market remains contested rather than settled.
SOFTWARE & BUYING
CUDA remains a draw
Reports describe continued use of NVIDIA hardware alongside domestic chips. Software support, availability, cost, and deployment reliability matter alongside speed.
05 / WHAT EVIDENCE COMES NEXT
From a headline number to a useful comparison
Independent, repeatable testing would show whether the reported advantages hold beyond the demonstration workloads.
Disclose the setup
Identify hardware configurations, software stacks, and model versions.
Repeat the tests
Publish methods and repeatability across the same workloads.
Expand workload coverage
Compare inference, training, image, video, throughput, and latency.
Show deployment facts
Report pricing, availability, power use, and software support.
THE TAKEAWAY
Competitive signals, not a verdict
Iluvatar has presented encouraging, workload-specific comparisons, including a reported 5% external throughput edge. Unclear test details, a company-run demo, and the H100 training comparison leave the scale of any advantage unresolved. Iluvatar’s stated goal to leapfrog NVIDIA’s Rubin platform within two years is a company target, not an independently verified forecast.
The evidence so far makes Tiangai 300 worth watching. It does not prove that it is faster than H200 across workloads or ready to replace established systems.
ASSESSMENT · REPORTED RESULTS REMAIN PRELIMINARY
A Small Lead in a Strategic Market
The results matter because Chinese technology companies face limits on access to advanced overseas accelerators, while domestic suppliers are under pressure to offer alternatives. The H200 is NVIDIA’s most advanced AI GPU approved for export to China under current U.S. rules, according to the reporting. A locally produced chip with competitive performance could give Chinese buyers another option for AI inference and training.
Still, the numbers presented so far suggest a possible narrow advantage in particular tests, not a replacement for NVIDIA across the market. The reported 1.05-times throughput result is a modest margin, and performance figures alone do not establish how a system performs across different models, software environments, costs, or large-scale deployments. NVIDIA’s CUDA software stack remains a major draw for companies, and reports indicate that local firms continue to use NVIDIA products alongside domestic chips.
That distinction is important for customers evaluating supply and performance. Even if Tiangai 300 delivers competitive results in selected workloads, buyers also need reliable software support, availability, and results that can be reproduced outside a vendor demonstration. The reported comparisons make the chip worth tracking, but do not settle whether it can displace established systems in routine use.
Products Worth Considering
Logitech MX Brio Ultra HD 4K Webcam, Graphite
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
DJI Osmo Mobile 7P Phone Gimbal with Mic Mini Transmitter, Infinity Black
As an affiliate, we earn on qualifying purchases.
China’s Push for Domestic Accelerators
Iluvatar CoreX unveiled Tiangai 300 in July and is one of several Chinese companies developing general-purpose GPUs. Other domestic competitors named in the reporting include Biren, MetaX, Enflame, and Moore Threads. In 2024, Iluvatar held 0.3% of China’s domestic general-purpose GPU market, a figure that underscores the gap between its current market position and its stated ambitions.
Chinese companies have not stopped buying or using NVIDIA hardware. The reporting describes CUDA as a major reason NVIDIA remains attractive, even as U.S. export controls restrict sales of some advanced accelerators. It also says that ByteDance, Alibaba, and Tencent have preferred NVIDIA’s RTX PRO 5500 over Huawei’s Ascend 950PR for inference. Those examples indicate continuing competition between domestic chips and NVIDIA products rather than a settled shift to one supplier.
Iluvatar was listed in Hong Kong earlier this year and set out a January roadmap that aims to leapfrog NVIDIA’s Rubin platform within two years. That is a company target, not an independently verified prediction. The Tiangai 300 results arrive as Chinese chipmakers seek to show that locally developed hardware can handle demanding AI tasks despite constraints on access to foreign technology.
“The company says Tiangai 300 can match and, in some cases, beat NVIDIA’s H200 in real-world AI workloads.”
— Iluvatar CoreX
Testing Conditions Remain Unclear
Most of the detailed Tiangai 300 versus H200 figures came from a company-run demonstration shown to mainland Chinese media. The reporting does not specify all test settings, system configurations, software versions, or repeatability data needed to independently evaluate the results. The separate INESA figure offers an external point of comparison, but the reported 5% margin is not enough by itself to establish a broad performance advantage.
The image-generation result also includes a trade-off: Tiangai 300 recorded a higher reported image rate and completed every task, while the H200’s completion rate was lower, but Tiangai 300 took 0.64 seconds longer to complete the task. The available account does not explain how that timing relates to the other reported metrics. It also provides no detailed text-to-video results, and the training comparison uses the H100 rather than the H200.
It remains unclear whether the reported advantages will hold across independent testing, a wider range of AI models, and production deployments. No detailed pricing, availability, power-consumption comparison, or independent replication was included in the reported results. Until those details are available, the claims should be treated as workload-specific and preliminary rather than as proof of general superiority.
Independent Results Will Matter
The next useful evidence would be repeatable third-party testing that identifies the workloads, hardware configurations, software stacks, and measurement methods used for both Tiangai 300 and the H200. Broader comparisons could show whether the reported throughput and latency results persist beyond the demonstration workloads and typical deployment conditions described so far.
Customers and industry observers will also be watching for details about Tiangai 300’s availability, pricing, and software support, as well as whether Iluvatar can expand beyond its reported market share. The company’s two-year roadmap sets an ambition to challenge NVIDIA’s Rubin platform, but that goal remains unverified. For now, the confirmed development is that Iluvatar has presented promising, limited comparisons; the scale and commercial significance of any advantage are still developing.
Source: Wccftech
Key Questions
Did Tiangai 300 prove faster than NVIDIA’s H200 overall?
No. Iluvatar reported wins in selected tests, and Shanghai INESA reportedly measured 1.05 times H200 throughput under typical deployment conditions. The available results do not establish that Tiangai 300 is faster across workloads generally.
What happened in the text-to-image comparison?
Iluvatar reported that Tiangai 300 completed every task at 0.101 images per second. The H200 completed 87.5% of tasks at 0.086 images per second. Tiangai 300 took 0.64 seconds longer to complete the task, according to the reported comparison.
Was the Tiangai 300 tested independently?
Most figures came from an Iluvatar demonstration. A separate assessment by Shanghai INESA reportedly found a 5% throughput advantage over the H200, but the available reporting does not provide detailed methods or independent replication results.
How does the training result compare with NVIDIA hardware?
Tiangai 300 reportedly achieved 1.33 times the effective throughput of the H100 while fine-tuning Alibaba’s Qwen3-32B. This result compares the chip with the H100, not the newer H200.
What is still unknown about Tiangai 300?
Detailed test settings, performance across more models, independent repeatability, pricing, availability, and power use were not provided in the reported results. Those details are needed to judge the chip’s broader value to customers.
Source: Wccftech
