Iluvatar CoreX’s New GPU Matches Nvidia H200 in Inference Tests
沐曦CoreX新款GPU在推理测试中性能追平英伟达H200
Building an Innovation Hub 打造创新中心
China Shaping the Future 中国塑造未来
by Zhu Shenshen 朱深深
[Tech] [科技]
Share Article: 分享文章：
Shanghai-based AI chip developer Iluvatar Corex publicly demonstrated its new Tiangai 300 GPU, revealing test results that match and slightly surpass those of Nvidia’s H200 chip in real-world AI inference tasks.总部位于上海的人工智能芯片研发公司壁仞科技公开展示了其新款天垓300图形处理器，测试结果显示，该产品在实际人工智能推理任务中的表现与英伟达H200芯片持平，且略有超越。
In tests attended by media in Shanghai, the Tiangai 300 achieved 1.05 times the single-card end-to-end throughput of the Nvidia H200 under standard disaggregated deployment settings.在上海有媒体出席的测试中，在标准解耦部署配置下，天垓300的单卡端到端吞吐量达到了英伟达H200的1.05倍。
Additionally, the chip outperformed the H200 in complex text-to-image and text-to-video multimodal workloads using major LLMs, including Zhipu AI’s GLM-5.3.此外，在使用包括智谱 AI 的 GLM-5.3 在内的主流大模型进行复杂文本转图像和文本转视频多模态任务时，该芯片的性能也优于 H200。
The Nvidia H200 currently represents the benchmark for high-performance AI computing approved for export to the Chinese mainland under US trade controls.英伟达H200目前是在美国贸易管制下获批出口至中国大陆的高性能人工智能计算设备的标杆产品。
Iluvatar Corex, which started listing in Hong Kong in January, is regarded as one of China’s “AI chip dragons,” offering the equivalent of Nvidia’s chips.1月在香港上市的壁仞科技（Iluvatar Corex）被视为中国的“AI芯片龙头”之一，其推出的芯片性能可与英伟达的芯片相媲美。
Analysts say the Tiangai 300’s performance indicates domestic Chinese GPU makers are transitioning from early-stage viability toward high-efficiency, enterprise-grade production capable of supporting large-scale commercial AI deployments.分析师表示，天垓300的性能表明中国国产GPU厂商正从早期可用阶段，向能够支撑大规模商业AI部署的高效能企业级生产阶段转型。
Credit: Ti Gong图片来源：Ti Gong
Caption: Testing Tiangai 300 and H200 in Shanghai.说明：在上海测试天改300与H200。
Editor: Yao Minji 编辑：姚敏吉
