Workflow
DeepSeek昨天悄悄扔的炸弹,今天爆了
Hu Xiu·2025-08-22 10:12

Core Viewpoint - The release of DeepSeek's V3.1 has sparked significant excitement in the A-share market, indicating a potential breakthrough for domestic chip manufacturers in the computing power sector [3][4][17]. Industry Impact - The A-share market saw over 2800 stocks rise, with computing power stocks and chip manufacturers experiencing substantial gains, highlighting a collective optimism in the sector [3][4]. - DeepSeek's statement is viewed as a signal of a "technical shift" and a rare moment of resonance within the domestic computing power industry [4][13]. Technical Details - The "UE8M0 FP8" format introduced by DeepSeek is designed to enhance the performance of AI models by compressing data to 8 bits, which reduces memory bandwidth pressure during training and inference [5][11]. - The UE8M0 format allows for efficient scaling of data blocks, significantly improving the dynamic range and reducing the risk of information loss during processing [10][12]. - The architecture of existing domestic AI accelerators is evolving to support the MX format, which is essential for overcoming memory bandwidth limitations [12][14]. Competitive Landscape - The release of UE8M0 FP8 signifies that domestic chip manufacturers have achieved a comprehensive integration of floating-point formats, compiler optimizations, and training framework adaptations [14]. - Major domestic chip companies, including Moore Threads and Cambrian, are already in contact with DeepSeek and are positioned to benefit from this new technology [15][16]. Future Outlook - DeepSeek's initiative is seen as a step towards reducing reliance on foreign technology, potentially redefining the efficiency of large model computing power and enhancing the international standing of domestic chip manufacturers [17].