Global GPU Computing News【20250823】

1. China's Intelligent Computing Scale Reaches 788 EFLOPS: The Ministry of Industry and Information Technology (MIIT) announced at the 2025 China Computing Conference that as of the end of June, China's intelligent computing scale reached 788 EFLOPS, with over 10.85 million standard racks in national computing centers, and the average PUE energy consumption metric for data centers dropped to 1.42.

2. MIIT Plans GPU Chip Core Technology Breakthroughs: MIIT Vice Minister Xiong Jijun stated plans to promote the systematic development of computing networks, focusing on breakthroughs in core technologies like GPU chips, and to launch a "unveiling the list" initiative for foundational computing strength, guiding the orderly construction of intelligent computing facilities.

3. Domestic GPU Sector Market Cap Exceeds Trillion Yuan: Influenced by news such as the rumored discontinuation of Nvidia's H20, shares of Chinese GPU companies like Cambricon and Hygon Information Technology rose. The combined market cap of related companies surpassed the trillion yuan mark on August 22nd, reflecting market confidence in domestic substitution.

4. DeepSeek V3.1's FP8 Precision Boosts Domestic Chips: DeepSeek released the DeepSeek-V3.1 model, adopting the UE8M0 FP8 precision format. This format is optimized for next-generation domestic chips and is expected to enhance the computational efficiency and performance of domestic AI chips in training and inference tasks.

5. "China Computing Platform" Connects Ten Provinces: The national computing power scheduling platform has connected sub-platforms in 10 provinces and municipalities, including Shanxi, Shanghai, Jiangsu, and Zhejiang, with over 1,000 registered enterprises, aiming to promote the coordination and sharing of computing resources.

6. Nvidia Introduces Spectrum-XGS Ethernet Technology: This technology focuses on "scale-across" capability, connecting data centers distributed globally into a unified, gigawatt-scale AI factory to optimize the efficiency and performance of large-scale AI training.

7. Rumors of Nvidia H20 Chip Discontinuation: Reports suggest Nvidia has instructed suppliers to pause production of the H20 AI accelerator chip designed for the Chinese market. Industry analysis attributes this to potential supply chain adjustments, security concerns, and the acceleration of domestic substitution.

8. FP8 Low-Precision Computing Emerges as Industry Trend: Nvidia published a paper demonstrating that its MXFP8 data format on Blackwell GPUs achieved nearly double the training efficiency compared to BF16 with almost no loss of precision, indicating wider adoption of FP8 in AI training and inference.

9. US Reportedly Attaches Conditions to Nvidia H20 Exports: According to reports, alongside approval for exporting the H20 chip to China, Nvidia is required to remit 15% of the sales revenue to the US government, increasing the chip's cost and market uncertainty.

10. CoreWeave Among First to Deploy Spectrum-XGS: The large-scale cloud provider CoreWeave will be among the first to deploy Nvidia's Spectrum-XGS Ethernet technology to connect its data centers and enhance its AI service capabilities.

Summary

This week's GPU computing news primarily focused on:

• Domestically (China): Continued growth in computing scale, strong policy support for core technology independence, accelerated domestic substitution, and industry ecosystem integration.

• Internationally: Industry giants like Nvidia pushing the boundaries of computing connectivity and efficiency with new technologies, while geopolitical factors continue to influence the market landscape.