The reported Tiangai 300 and NVIDIA H200 results point in different directions: Tiangai 300 had higher image throughput and completion in one text-to-image test, but took 0.64 seconds longer to finish the task. A separate test attributed to Shanghai INESA reported 1.05× H200 throughput under conditions described as typical deployment. These figures support workload-specific comparisons, not a general performance ranking.
Reported results side by side
Wccftech reported the text-to-image figures on October 31, 2026. The report date is not an established test date. It gave these results for one test and separately attributed the 1.05× throughput result to Shanghai INESA:
| Reported measure | Tiangai 300 | NVIDIA H200 | Test context |
| Text-to-image throughput | 0.101 images per second | 0.086 images per second | One reported text-to-image test |
| Text-to-image completion rate | 100% | 87.5% | One reported text-to-image test |
| Throughput per accelerator | 1.05× H200 throughput | 1.00× baseline | Separate Shanghai INESA result, under conditions described as typical deployment |
The 1.05× result means Tiangai 300 had 5% higher throughput than H200 in that reported test. Wccftech described the conditions as typical deployment, but its account did not give the workload details or methodology. That result does not turn the separate text-to-image figures into a single benchmark.
Higher throughput did not mean a shorter task time
Wccftech also reported that Tiangai 300 took 0.64 seconds longer than H200 to complete the text-to-image task. Throughput measures how much work is completed over time; task completion time measures how long that task took. The reported figures therefore favor Tiangai 300 on throughput and completion rate, while H200 finished the task sooner.
What the media demonstration covered
Iluvatar CoreX said Tiangai 300 matched or outperformed H200 in a media demonstration of text-to-image and text-to-video generation. The South China Morning Post reported on October 29, 2026, that the demonstration had taken place on the preceding Wednesday and included models such as Z.ai’s GLM-5.3. That account did not provide benchmark figures; the numerical results above come from Wccftech’s later report.
