Iluvatar CoreX Says Tiangai 300 GPU Matches Nvidia H200 in First Live Shanghai Demo

Iluvatar CoreX, the Shanghai chip designer formally known as Shanghai Tianshu Zhixin Integrated Circuit, used its 2026 media open day on October 28 to run its flagship Tiangai 300 general-purpose GPU on real hardware in front of reporters for the first time. The company said the part is the first domestic general-purpose GPU whose measured performance reaches the level of Nvidia's H200, citing a third-party benchmark that put per-card end-to-end throughput at 1.05 times the American accelerator.

The claim arrives in a market shaped by export controls. Nvidia's Blackwell-generation parts are barred from sale to Chinese customers, which leaves the H200 as the most advanced Nvidia AI processor cleared for export to China under current US trade rules. For buyers who cannot obtain newer Nvidia silicon, the H200 is the practical yardstick, and any domestic chip that credibly approaches it changes how operators weigh price, supply risk and software.

Iluvatar CoreX is not a newcomer to this argument. The Tiangai 300 was formally launched at the World Artificial Intelligence Conference in July 2026, built on the company's Tianshu architecture with a new generation of Tensor Cores and 144GB of on-card memory.

Iluvatar CoreX presented results from testing carried out by Shanghai INESA in July 2026 and paired those figures with a live demonstration using frontier models. The test conditions have not been published, the H200 remains a part designed by a company with a long software lead, and the numbers are vendor-disclosed.

Key Facts

The South China Morning Post reported on October 29 that Iluvatar CoreX demonstrated the Tiangai 300 outperforming the H200 in complex multimodal work at a live event for mainland Chinese media on Wednesday, October 28. The demonstration covered text-to-image and text-to-video generation running on frontier models including Z.ai's GLM-5.3, the newspaper said, attributing parts of its account to Chinese tech outlet Xinzhixun. IT Home reported on October 28 that the media day was the first public showing of the Tiangai 300 running on physical hardware. The outlet quoted the company's claim that a third-party measurement, converted to a per-card basis, reached 1.05 times the H200's overall throughput. It also flagged the limitation plainly: the test method, model loads and workload conditions had not been released, so the result remains a vendor-disclosed figure for a particular configuration rather than a reproducible conclusion.

Zhidx reported on October 28 that the Tiangai 300 uses the Tianshu architecture, carries 144GB of memory, natively supports scalar, vector and tensor compute modes, and covers FP4, FP8, FP16 and BF16 precisions. In the company's performance laboratory, attention throughput at a 64K context reached 1.05 times the H200, MoE throughput with GLM-5.3 at 64K reached 1.15 times the H200, and agent first-token latency across 8K to 128K contexts fell by roughly 10 percent to 30 percent. Online recommendation workloads saw 21 percent higher throughput and 26 percent lower latency in lightweight compute.

Training results were included as well. On an eight-card, FP8 training demonstration using Qwen3-32B, the Tiangai 300 reached 1.33 times the per-card effective throughput of Nvidia's H100. Iluvatar CoreX also offered a conditional engineering estimate: with 1,024 cards applied to reinforcement learning and 40 billion activated parameters as the central scenario, a model with roughly 1 trillion total parameters could complete a task generating 50 billion tokens in about 26 days.

Yonhap News Agency reported on October 30 that in the text-to-image task the Tiangai 300 finished 100 percent of jobs at 0.101 images per second, against an 87.5 percent completion rate at 0.086 images per second for the H200. A countervailing detail: the Tiangai 300 took 0.64 seconds longer than the H200 to complete each task. The report tied the demonstration to the July INESA test, which found 5 percent higher throughput under ordinary deployment conditions and 33 percent higher effective throughput than the H100.

The IXL144 supernode, first shown at the China International Industry Fair on October 12 and 13, packs 144 GPUs and 18 CPUs into a single cabinet with roughly 10.1 terabytes of memory. Iluvatar CoreX says the configuration cuts the cost per 1 million valid tokens by 40 percent compared with mainstream international alternatives, and describes it as the highest-density single-cabinet GPU supernode in the industry. A heterogeneous AFD split deployment, which divides attention and feed-forward stages across GPU and XPU units, reduced inference latency by 60 percent, the company said.

Analysis

What this really means is that the competitive question for Chinese AI accelerators has moved. The test was once whether a domestic part could run a demanding model at all; the test now being staged in public is whether it can run that model at comparable throughput and cost. A chip that works but trails badly is a policy hedge, useful for compliance and continuity but not for a procurement fight. A chip that works at rough parity becomes a purchasing decision.

The evidence is directional rather than conclusive, and that gap is the story's centre of gravity. Throughput figures depend heavily on batch size, sequence length, quantization, scheduler behaviour and comparison-system tuning, and the H200 side of this comparison was measured under conditions the public has not seen. The image-generation result shows the texture: a higher completion rate and a higher images-per-second figure sit alongside a slower time to finish each job, which points to an advantage in aggregate throughput under sustained load rather than in single-request responsiveness.

The bigger picture here is that the unit of competition has become the cabinet, not the card. The 1.05x per-card number is the headline, but the IXL144 figure of 144 GPUs, 18 CPUs and about 10.1 terabytes of memory in one rack, with a claimed 40 percent reduction in cost per million valid tokens, is the number a procurement committee can use. Token economics decide whether an inference cluster pays for itself, and a 40 percent cost reduction claim, if it survives scrutiny, matters more to a buyer than a five percent throughput edge.

Software remains the variable the benchmarks do not measure. Nvidia's position rests as much on its programming model and the years of tuned kernels built on it as on transistor-level performance, and the materials released this week say nothing about compiler maturity or framework coverage. The demonstration ran GLM-5.3, a domestic frontier model: a model tuned for domestic silicon can look considerably better than an arbitrary production workload dropped onto the same hardware. None of this invalidates the result, but it bounds what the result can prove.

Why It Matters

The policy backdrop gives a benchmark a second life. Washington has eased restrictions on H200 sales to China, yet Beijing has been discouraging domestic buyers from leaning on foreign chips, according to Yonhap's account. That creates a strange market in which the best available Nvidia part is legally obtainable but politically discouraged, and a credible local substitute has a captive opening. Iluvatar CoreX is not only selling throughput; it is selling an option that does not require anyone to ask for permission.

The vendor landscape adds pressure. Yonhap described the sector as grouping Enflame, Moore Threads, Biren Technology and MetaX as the so-called four dragons, with Iluvatar CoreX added to form a group of five. In a field that crowded, a public claim of H200 parity is as much a positioning move against domestic rivals as against Nvidia, and being first to say "measured at H200 level" carries weight with customers and investors.

For the broader hardware market, the export-control strategy is being tested at the exact performance tier it was designed to protect. If a Chinese general-purpose GPU genuinely reaches H200 throughput per card, restricting Blackwell loses some of its bite, because the fallback tier stops being a generation or two behind and starts being competitive on the metric buyers care about most: cost per token served. That is a slower kind of gap closing than a leapfrog, and it is the kind that compounds.

Next Up

The immediate test is reproducibility. The first thing customers and independent engineers will ask for is the methodology: model list, batch and sequence configuration, and the raw logs behind the 1.05x figure. Until that material appears, the claim stands as a vendor result from a July test by Shanghai INESA, restated at an October media day. A target of cutting intra-domain transfer latency from 3 to 5 microseconds down to 1 microsecond will be read as evidence that the roadmap is holding its dates.

The product line beyond the data centre cabinet is already queued. Iluvatar CoreX used the same event to present the Tongyang TY1200 edge compute terminal, specified at 300 TOPS of dense compute for AI PCs and embodied intelligence, saying it outperformed Nvidia's Thor T5000 in measured tests and achieved GPU-accelerated physics simulation based on MuJoCo-Warp on a domestic edge product. It also previewed the TY2000, with a nominal 600T of compute and support for multi-machine interconnection. Whether the Tiangai 300 claim survives scrutiny will determine how much weight the edge claims carry.